Fueled by the Large Language Models (LLMs) wave, Large Visual-Language Models (LVLMs) have emerged as a pivotal advancement, bridging the gap between image and text. However, video making it challenging for LVLMs to perform adequately due to the complexity of the relationship between language and spatial-temporal data structure. Recent Large Video-Language Models (LVidLMs) align feature of static visual data like image into latent space of language feature, by general multi-modal tasks to leverage abilities of LLMs sufficiently. In this paper, we explore fine-grained alignment approach via object trajectory for different modalities across both spatial and temporal dimensions simultaneously. Thus, we propose a novel LVidLM by trajectory-guided Pixel-Temporal Alignment, dubbed PiTe, that exhibits promising applicable model property. To achieve fine-grained video-language alignment, we curate a multi-modal pre-training dataset PiTe-143k, the dataset provision of moving trajectories in pixel level for all individual objects, that appear and mention in the video and caption both, by our automatic annotation pipeline. Meanwhile, PiTe demonstrates astounding capabilities on myriad video-related multi-modal tasks through beat the state-of-the-art methods by a large margin.

Self-supervised learning (SSL) has proven effective in learning high-quality representations for various downstream tasks, with a primary focus on semantic tasks. However, its application in geometric tasks remains underexplored, partially due to the absence of a standardized evaluation method for geometric representations. To address this gap, we introduce a novel pose-estimation benchmark for assessing SSL geometric representations, which demands training without semantic or pose labels and achieving proficiency in both semantic and geometric downstream tasks. On this benchmark, we study enhancing SSL geometric representations without sacrificing semantic classification accuracy. We find that leveraging mid-layer representations improves pose-estimation performance by 10-20%. Further, we introduce an unsupervised trajectory-regularization loss, which improves performance by an additional 4% and improves generalization ability on out-of-distribution data. We hope the proposed benchmark and methods offer new insights and improvements in self-supervised geometric representation learning.

While recent vision-and-language models (VLMs) like CLIP are a powerful tool for analyzing text and images in a shared semantic space, they do not explicitly model the hierarchical nature of the set of texts which may describe an image. Conversely, existing multimodal hierarchical representation learning methods require costly training from scratch, failing to leverage the knowledge encoded by state-of-the-art multimodal foundation models. In this work, we study the knowledge of existing foundation models, finding that they exhibit emergent understanding of visual-semantic hierarchies despite not being directly trained for this purpose. We propose the Radial Embedding (RE) framework for probing and optimizing hierarchical understanding, and contribute the HierarCaps dataset, a benchmark facilitating the study of hierarchical knowledge in image--text representations, constructed automatically via large language models. Our results show that foundation VLMs exhibit zero-shot hierarchical understanding, surpassing the performance of prior models explicitly designed for this purpose. Furthermore, we show that foundation models may be better aligned to hierarchical reasoning via a text-only fine-tuning phase, while retaining pretraining knowledge. We will release our data, code, and trained models.

Multimodal models have gained increasing popularity recently. Many works have been proposed to learn the representations for different modalities. The representation can learn shared information from these domains, leading to increased and coherent joint and cross-generation. However, these works mainly considered standard Gaussian or Laplacian as their prior distribution. It can be challenging for the uni-modal and non-informative distribution to capture all the information from multiple data types. Meanwhile, energy-based models (EBM) have shown their effectiveness in multiple tasks due to their expressiveness and flexibility. But its capacity has yet to be discovered for the multimodal generative models. In this paper, we propose a novel framework to train multimodal latent generative models together with the energy-based models. The proposed method can lead to more expressive and informative prior which can better capture the information within multiple modalities. Our experiments showed that our model is effective and can increase generation coherence and latent classification for different multimodal datasets.

The increasing availability of multi-sensor data sparks interest in multimodal self-supervised learning. However, most existing approaches learn only common representations across modalities while ignoring intra-modal training and modality-unique representations. We propose Decoupling Common and Unique Representations (DeCUR), a simple yet effective method for multimodal self-supervised learning. By distinguishing inter- and intra-modal embeddings through multimodal redundancy reduction, DeCUR can integrate complementary information across different modalities. Meanwhile, a simple residual deformable attention is introduced to help the model focus on modality-informative features. We evaluate DeCUR in three common multimodal scenarios ( radar-optical, RGB-elevation, and RGB-depth), and demonstrate its consistent and significant improvement for both multimodal and modality-missing settings. With thorough experiments and comprehensive analysis, we hope this work can provide insights and raise more interest in researching the hidden relationships of multimodal representations.

Vision Transformer models trained on large-scale datasets, although effective, often exhibit artifacts in the patch token they extract. While such defects can be alleviated by re-training the entire model with additional classification tokens, the underlying reasons for the presence of these tokens remain unclear. In this paper, we conduct a thorough investigation of this phenomenon, combining theoretical analysis with empirical observations. Our findings reveal that these artifacts originate from the pre-trained network itself, specifically stemming from the leading left singular vector of the network's weights. Furthermore, to mitigate these defects, we propose a novel fine-tuning smooth regularization that rectifies structural deficiencies using only a small dataset, thereby avoiding the need for complete re-training. We validate our method on various downstream tasks, including unsupervised segmentation, classification, and supervised segmentation, demonstrating its effectiveness in improving model performance. Our code and checkpoints will be released.

We delve into a crucial yet often overlooked challenge inherent to Vision Transformers (ViTs): feature maps of these models exhibit grid-like artifacts, which hurt the performance of ViTs in downstream dense prediction tasks such as segmentation, depth prediction, and object discovery. We trace this fundamental issue down to the positional embeddings at the input stage. we propose a two-stage denoising approach, termed Denoising Vision Transformers (DVT). In the first stage, we separate the clean features from those contaminated by positional artifacts by enforcing cross-view feature consistency with neural fields on a per-image basis. This per-image optimization process extracts artifact-free features from raw ViT outputs, providing clean feature estimates for offline applications. In the second stage, we train a lightweight Transformer block to predict clean features from raw ViT outputs, leveraging the derived estimates of the clean features as supervision. Our DVT does not require re-training the existing pre-trained ViTs, and is immediately applicable to any Vision Transformer architecture. We evaluate our method on a variety of representative ViTs (DINO, DeiT-III, EVA02, CLIP, DINOv2, DINOv2-reg) and demonstrate that our DVT consistently and significantly improves existing state-of-the-art general-purpose models in semantic and geometric tasks across multiple datasets. We hope our study will …

Recent advancements in transformer-based light-weight object tracking have set new standards across various benchmarks due to their efficiency and effectiveness. Despite these achievements, most current trackers rely heavily on pre-existing object detection architectures without optimizing the backbone network to leverage the unique demands of object tracking. Addressing this gap, we introduce the Feature Extraction and Relation Modeling Tracker (FERMT) - a novel approach that significantly enhances tracking speed and accuracy. At the heart of FERMT is a strategic decomposition of the conventional attention mechanism into four distinct sub-modules within a one-stream tracker. This design stems from our insight that the initial layers of a tracking network should prioritize feature extraction, whereas the deeper layers should focus on relation modeling between objects. Consequently, we propose an innovative, light-weight backbone specifically tailored for object tracking. Our approach is validated through meticulous ablation studies, confirming the effectiveness of our architectural decisions. Furthermore, FERMT incorporates a Dual Attention Unit for feature pre-processing, which facilitates global feature interaction across channels and enriches feature representation with attention cues. Benchmarking on GOT-10k, FERMT achieves a groundbreaking Average Overlap (AO) score of 69.6%, outperforming the leading real-time trackers by 5.6% in accuracy while boasting a 54% improvement in …

Predicting realistic ground views from satellite imagery in urban scenes is a challenging task due to the significant view gaps between satellite and ground-view images. We propose a novel pipeline to tackle this challenge, by generating geospecifc views that maximally respect the weak geometry and texture from multi-view satellite images. Different from existing approaches that hallucinate images from cues such as partial semantics or geometry from overhead satellite images, our method directly predicts ground-view images at geolocation by using a comprehensive set of information from the satellite image, resulting in ground-level images with a resolution boost at a factor of ten or more. We leverage a novel building refinement method to reduce geometric distortions in satellite data at ground level, which ensures the creation of accurate conditions for view synthesis using diffusion networks. Moreover, we proposed a novel geospecific prior, which prompts distribution learning of diffusion models to respect image samples that are closer to the geolocation of the predicted images. We demonstrate our pipeline is the first to generate close-to-real and geospecific ground views merely based on satellite images. Codes and data will be shared.

6-DoF grasp detection has been a fundamental and challenging problem in robotic vision. While previous works have focused on ensuring grasp stability, they often do not consider human intention conveyed through natural language, hindering effective collaboration between robots and users in complex 3D environments. In this paper, we present a new approach for language-driven 6-DoF grasp detection in cluttered point clouds. We first introduce Grasp-Anything-6D, a large-scale dataset for the language-driven 6-DoF grasp detection task with 1M point cloud scenes and more than 200M language-associated 3D grasp poses. We further introduce a novel diffusion model that incorporates a new negative prompt guidance learning strategy. The proposed negative prompt strategy directs the detection process toward the desired object while steering away from unwanted ones given the language input. Our method enables an end-to-end framework where humans can command the robot to grasp desired objects in a cluttered scene using natural language. Intensive experimental results show the effectiveness of our method in both benchmarking experiments and real-world scenarios, surpassing other baselines. In addition, we demonstrate the practicality of our approach in real-world robotic applications.

We aim to discover manipulation concepts embedded in the unannotated demonstrations, which are recognized as key physical states. The discovered concepts can facilitate training manipulation policies and promote generalization. Current methods relying on multimodal foundation models for deriving key states usually lack accuracy and semantic consistency due to limited multimodal robot data. In contrast, we introduce an information-theoretic criterion to characterize the regularities that signify a set of physical states. We also develop a framework that trains a concept discovery network using this criterion, thus bypassing the dependence on human semantics and alleviating costly human labeling. The proposed criterion is based on the observation that key states, which deserve to be conceptualized, often admit more physical constraints than non-key states. This phenomenon can be formalized as maximizing the mutual information between the putative key state and its preceding state, i.e., Maximal Mutual Information (MaxMI). By employing MaxMI, the trained key state localization network can accurately identify states of sufficient physical significance, exhibiting reasonable semantic compatibility with human perception. Furthermore, the proposed framework produces key states that lead to concept-guided manipulation policies with higher success rates and better generalization in various robotic tasks compared to the baselines, verifying the effectiveness of …

Collaborative perception has received widespread attention recently since it enhances the perception ability of autonomous vehicles via inter-agent information sharing. However, the performance of existing systems is hindered by the unavoidable collaboration noises, which induce feature-level spatial misalignment over the collaborator-shared information. In this paper, we propose a model-agnostic and lightweight plugin to mitigate the feature-level misalignment issue, called dynamic feature alignment (NEAT). The merits of the NEAT plugin are threefold. First, we introduce an importance-guided query proposal to predict potential foreground regions with space-channel semantics and exclude environmental redundancies. On this basis, a deformable feature alignment is presented to explicitly align the collaborator-shared features through query-aware spatial associations, aggregating multi-grained visual clues with corrective mismatch properties. Ultimately, we perform a region cross-attention reinforcement to facilitate aligned representation diffusion and achieve global feature semantic enhancement. NEAT can be readily inserted into existing collaborative perception procedures and significantly improves the robustness of vanilla baselines against pose errors and transmission delay. Extensive experiments on four collaborative 3D object detection datasets under noisy settings confirm that NEAT provides consistent gains for most methods with distinct structures.

With the comprehensive research conducted on various face analysis tasks, there is a growing interest among researchers to develop a unified approach to face perception. Existing methods mainly discuss unified representation and training, which lack task extensibility and application efficiency. To tackle this issue, we focus on the unified model structure, exploring a face generalist model. As an intuitive design, Naive Faceptor enables tasks with the same output shape and granularity to share the structural design of the standardized output head, achieving improved task extensibility. Furthermore, Faceptor is proposed to adopt a well-designed single-encoder dual-decoder architecture, allowing task-specific queries to represent new-coming semantics. This design enhances the unification of model structure while improving application efficiency in terms of storage overhead. Additionally, we introduce Layer-Attention into Faceptor, enabling the model to adaptively select features from optimal layers to perform the desired tasks. Through joint training on 13 face perception datasets, Faceptor achieves exceptional performance in facial landmark localization, face parsing, age estimation, expression recognition, binary attribute classification, and face recognition, achieving or surpassing specialized methods in most tasks. Our training framework can also be applied to auxiliary supervised learning, significantly improving performance in data-sparse tasks such as age estimation and expression …

Robustness is the most important property of watermarking schemes. In practice, the watermarking mechanism shall be robust to both geometric and non-geometric distortions. In deep learning-based watermarking frameworks, robustness can be ensured by end-to-end training with different noise layers. However, most of the current CNN-based watermarking frameworks, even trained with targeted distortions, cannot well adapt to geometric distortions due to the architectural design. Since the traditional convolutional layer's position structure is relatively fixed, it lacks the flexibility to capture the influence of geometric distortion, making it difficult to train for corresponding robustness. To address such limitations, we propose a Swin Transformer and Deformable Convolutional Network (DCN)-based watermark model backbone. The attention mechanism and the deformable convolutional window effectively improve the feature processing flexibility, greatly enhancing the robustness, especially for geometric distortions. Besides, for non-geometric distortions, aiming at improving the generalizability for more distortions, we also provide a distortion-style-ensembled noise layer, including an image encoder, an image decoder, and distortion-style layers that can effectively simulate styles of different kinds of distortions. In the final watermark model training stage, we can simply train with our proposed noise layer for overall robustness. Extensive experiments illustrate that compared to existing state-of-the-art (SOTA) works, with …

The generation of large-scale urban layouts has garnered substantial interest across various disciplines. Prior methods have utilized procedural generation requiring manual rule coding or deep learning needing abundant data. However, prior approaches have not considered the context-sensitive nature of urban layout generation. Our approach addresses this gap by leveraging a canonical graph representation for the entire city, which facilitates scalability and captures the multi-layer semantics inherent in urban layouts. We introduce a novel graph-based masked autoencoder (GMAE) for city-scale urban layout generation. The method encodes attributed buildings, city blocks, communities and cities into a unified graph structure, enabling self-supervised masked training for graph autoencoder. Additionally, we employ scheduled iterative sampling for 2.5D layout generation, prioritizing the generation of important city blocks and buildings. Our approach achieves good realism, semantic consistency, and correctness across the heterogeneous urban styles in 330 US cities. Codes and datasets are released at: https://github.com/Arking1995/COHO.

Segmenting and recognizing a diverse range of object parts is crucial in various computer vision and robotic applications. While object segmentation has made significant progress, part-level segmentation remains an under-explored issue. Part segmentation entails discerning complex boundaries between parts, and the scarcity of annotated data further complicates the task. To tackle this problem, in this paper, we propose a novel Weakly-supervised Part Segmentation (WPS) setting and an approach called WPS-SAM, built on the large-scale pre-trained vision foundation model, Segment Anything Model (SAM). WPS-SAM is an end-to-end framework designed to extract prompt tokens directly from images and perform pixel-level segmentation of part regions. During its training phase, it only utilizes weakly supervised labels in the form of bounding boxes or points. Extensive experiments demonstrate that, through exploiting the rich knowledge embedded in pre-trained foundation models, WPS-SAM outperforms other segmentation models trained with pixel-level strong annotations. Specifically, WPS-SAM achieves 68.93% mIOU and 79.53% mACC on the PartImageNet dataset, surpassing state-of-the-art fully supervised methods by approximately 4% in terms of mIOU.

Text-to-image diffusion models have shown remarkable success in synthesizing photo-realistic images. Apart from creative applications, can we use such models to synthesize samples that aid the few-shot training of discriminative models? In this work, we propose AlignDiff, a general framework for synthesizing training images and masks for few-shot segmentation. We identify two crucial misalignments that arise when utilizing pre-trained diffusion models in segmentation tasks, which need to be addressed to create realistic training samples and align the synthetic data distribution with the real training distribution: 1) instance-level misalignment, where generated samples of rare categories are often misaligned with target tasks) and 2) annotation-level misalignment, where diffusion models are limited to generating images without pixel-level annotations. AlignDiff overcomes both challenges by leveraging a few real samples to guide the generation, thus improving novel IoU over baseline methods in few-shot segmentation and generalized few-shot segmentation on Pascal-5i and COCO-20i by up to 80%. Notably, AlignDiff is capable of augmenting the learning of out-of-distribution uncommon categories on FSS-1000, while naive diffusion model generates samples that diminish segmentation performance.

The Segment Anything Model (SAM) has demonstrated remarkable zero-shot capability and flexible geometric prompting in general image segmentation. However, it often struggles in domains that are either sparsely represented or lie outside its training distribution, such as aerial, medical, and non-RGB images. Recent efforts have predominantly focused on adapting SAM to these domains using fully supervised methods, which necessitate large amounts of annotated training data and pose practical challenges in data collection. This paper presents CAT-SAM, a ConditionAl Tuning network that explores few-shot adaptation of SAM toward various challenging downstream domains in a data-efficient manner. The core design is a prompt bridge structure that enables decoder-conditioned joint tuning of the heavyweight image encoder and the lightweight mask decoder. The bridging maps the domain-specific features of the mask decoder to the image encoder, fostering synergic adaptation of both components with mutual benefits with few-shot target samples only, ultimately leading to superior segmentation in various downstream tasks. We develop two CAT-SAM variants that adopt two tuning strategies for the image encoder: one injecting learnable prompt tokens in the input space and the other inserting lightweight adapter networks. Extensive experiments over 11 downstream tasks show that CAT-SAM achieves superior segmentation consistently even under …

Pre-trained vision-language models, e.g., CLIP, have been increasingly used to address the challenging Open-Vocabulary Segmentation (OVS) task, benefiting from their well-aligned vision-text embedding space. Typical solutions involve either freezing CLIP during training to unilaterally maintain its zero-shot capability, or fine-tuning CLIP vision encoder to achieve perceptual sensitivity to local regions. However, few of them incorporate vision-text collaborative optimization. Based on this, we propose the Content-Dependent Transfer to adaptively enhance each text embedding by interacting with the input image, which presents a parameter-efficient way to optimize the text representation. Besides, we additionally introduce a Representation Compensation strategy, reviewing the original CLIP-V representation as compensation to maintain the zero-shot capability of CLIP. In this way, the vision and text representation of CLIP are optimized collaboratively, enhancing the alignment of the vision-text feature space. To the best of our knowledge, we are the first to establish the collaborative vision-text optimizing mechanism within the OVS field. Extensive experiments demonstrate our method achieves superior performance on popular OVS benchmarks. In open-vocabulary semantic segmentation, our method outperforms the previous state-of-the-art approaches by +0.5, +2.3, +3.4, +0.4 and +1.1 mIoU, respectively on A-847, A-150, PC-459, PC-59 and PAS-20. Furthermore, in a panoptic setting on the ADE20K dataset, …

Unsupervised Domain Adaptation (UDA) for semantic segmentation has been widely studied to exploit the label-rich source data to assist the segmentation of unlabeled samples on target domain. Despite these efforts, UDA performance remains far below that of fully-supervised model owing to the lack of target annotations. To this end, we propose an efficient superpixel-level active learning method for domain adaptive semantic segmentation to maximize segmentation performance by automatically querying a small number of superpixels for labeling. To conserve annotation resources, we propose a novel low-uncertainty superpixel fusion module which amalgamates superpixels possessing low-uncertainty features based on feature affinity and thereby ensuring high-quality fusion of superpixels. As for the acquisition strategy, our method takes into account two types of information-rich superpixels: large-size superpixels with substantial information content, and superpixels with the greatest value for domain adaptation learning. Further, we employ the cross-domain mixing and pseudo label with consistency regularization techniques respectively to address the domain shift and label noise problems. Extensive experimentation demonstrates that our proposed superpixel-level method utilizes a limited budget more efficiently than previous pixel-level techniques and surpasses state-of-the-art methods at 40x lower cost.

Delving into the realm of egocentric vision, the advancement of referring video object segmentation (RVOS) stands as pivotal in understanding human activities. However, existing RVOS task primarily relies on static attributes such as object names to segment target objects, posing challenges in distinguishing target objects from background objects and in identifying objects undergoing state changes. To address these problems, this work proposes a novel action-aware RVOS setting called ActionVOS, aiming at segmenting only active objects in egocentric videos using human actions as a key language prompt. This is because human actions precisely describe the behavior of humans, thereby helping to identify the objects truly involved in the interaction and to understand possible state changes. We also build a method tailored to work under this specific setting. Specifically, we develop an action-aware labeling module with an efficient action-guided focal loss. Such designs enable ActionVOS model to prioritize active objects with existing readily-available annotations. Experimental results on VISOR dataset reveal that ActionVOS significantly reduces the mis-segmentation of inactive objects, confirming that actions help the ActionVOS model understand objects' involvement. Further evaluations on VOST and VSCOS datasets show that the novel ActionVOS setting enhances segmentation performance when encountering challenging circumstances involving object state …

Image modality is not perfect as it often fails in certain conditions, e.g., night and fast motion. This significantly limits the robustness and versatility of existing multi-modal (i.e., Image+X) semantic segmentation methods when confronting modality absence or failure, as often occurred in real-world applications. Inspired by the open-world learning capability of multi-modal vision-language models (MVLMs), we explore a new direction in learning the modality-agnostic representation via knowledge distillation (KD) from MVLMs. Intuitively, we propose Any2Seg, a novel framework that can achieve robust segmentation from any combination of modalities in any visual conditions. Specifically, we first introduce a novel language-guided semantic correlation distillation (LSCD) module to transfer both inter-modal and intra-modal semantic knowledge in the embedding space from MVLMs, e.g., LanguageBind. This enables us to minimize the modality gap and alleviate semantic ambiguity to combine any modalities in any visual conditions. Then, we introduce a modality-agnostic feature fusion (MFF) module that reweights the multi-modal features based on the inter-modal correlation and selects the fine-grained feature. This way, our Any2Seg finally yields an optimal modality-agnostic representation. Extensive experiments on two benchmarks with four modalities demonstrate that Any2Seg achieves the state-of-the-art under the multi-modal setting (+3.54 mIoU) and excels in the challenging modality-incomplete …

Open-vocabulary segmentation is the task of segmenting anything that can be named in an image. Recently, large-scale vision-language modelling has led to significant advances in open-vocabulary segmentation, but at the cost of gargantuan and increasing training and annotation efforts. Hence, we ask if it is possible to use existing foundation models to synthesise on-demand efficient segmentation algorithms for specific class sets, making them applicable in an open-vocabulary setting without the need to collect further data, annotations or perform training. To that end, we present OVDiff, a novel method that leverages generative text-to-image diffusion models for unsupervised open-vocabulary segmentation. OVDiff synthesises support image sets for arbitrary textual categories, creating for each a set of prototypes representative of both the category and its surrounding context (background). It relies solely on pre-trained components and outputs the synthesised segmenter directly, without training. Our approach shows strong performance on a range of benchmarks, obtaining a lead of more than 5% over prior work on PASCAL VOC.

Since the introduction of NeRFs, considerable attention has been focused on improving their training and inference times, leading to the development of Fast-NeRFs models. Despite demonstrating impressive rendering speed and quality, the rapid convergence of such models poses challenges for further enhancing reconstruction quality. Common strategies to improve rendering quality involves augmenting model parameters or increasing the number of sampled points. However, these computationally intensive approaches encounter limitations in achieving significant quality enhancements. This study introduces a model-agnostic framework inspired by Sparsely-Gated Mixture of Experts to enhance rendering quality without escalating computational complexity. Our approach enables specialization in rendering different scene components by employing a mixture of experts with varying resolutions. We present a novel gate formulation designed to maximize expert capabilities and propose a resolution-based routing technique to effectively induce sparsity and decompose scenes. Our work significantly enhances reconstruction quality while maintaining competitive performance.

Image modality is not perfect as it often fails in certain conditions, e.g., night and fast motion. This significantly limits the robustness and versatility of existing multi-modal (i.e., Image+X) semantic segmentation methods when confronting modality absence or failure, as often occurred in real-world applications. Inspired by the open-world learning capability of multi-modal vision-language models (MVLMs), we explore a new direction in learning the modality-agnostic representation via knowledge distillation (KD) from MVLMs. Intuitively, we propose Any2Seg, a novel framework that can achieve robust segmentation from any combination of modalities in any visual conditions. Specifically, we first introduce a novel language-guided semantic correlation distillation (LSCD) module to transfer both inter-modal and intra-modal semantic knowledge in the embedding space from MVLMs, e.g., LanguageBind. This enables us to minimize the modality gap and alleviate semantic ambiguity to combine any modalities in any visual conditions. Then, we introduce a modality-agnostic feature fusion (MFF) module that reweights the multi-modal features based on the inter-modal correlation and selects the fine-grained feature. This way, our Any2Seg finally yields an optimal modality-agnostic representation. Extensive experiments on two benchmarks with four modalities demonstrate that Any2Seg achieves the state-of-the-art under the multi-modal setting (+3.54 mIoU) and excels in the challenging modality-incomplete …

Open-vocabulary segmentation is the task of segmenting anything that can be named in an image. Recently, large-scale vision-language modelling has led to significant advances in open-vocabulary segmentation, but at the cost of gargantuan and increasing training and annotation efforts. Hence, we ask if it is possible to use existing foundation models to synthesise on-demand efficient segmentation algorithms for specific class sets, making them applicable in an open-vocabulary setting without the need to collect further data, annotations or perform training. To that end, we present OVDiff, a novel method that leverages generative text-to-image diffusion models for unsupervised open-vocabulary segmentation. OVDiff synthesises support image sets for arbitrary textual categories, creating for each a set of prototypes representative of both the category and its surrounding context (background). It relies solely on pre-trained components and outputs the synthesised segmenter directly, without training. Our approach shows strong performance on a range of benchmarks, obtaining a lead of more than 5% over prior work on PASCAL VOC.

Pre-trained vision-language models, e.g., CLIP, have been increasingly used to address the challenging Open-Vocabulary Segmentation (OVS) task, benefiting from their well-aligned vision-text embedding space. Typical solutions involve either freezing CLIP during training to unilaterally maintain its zero-shot capability, or fine-tuning CLIP vision encoder to achieve perceptual sensitivity to local regions. However, few of them incorporate vision-text collaborative optimization. Based on this, we propose the Content-Dependent Transfer to adaptively enhance each text embedding by interacting with the input image, which presents a parameter-efficient way to optimize the text representation. Besides, we additionally introduce a Representation Compensation strategy, reviewing the original CLIP-V representation as compensation to maintain the zero-shot capability of CLIP. In this way, the vision and text representation of CLIP are optimized collaboratively, enhancing the alignment of the vision-text feature space. To the best of our knowledge, we are the first to establish the collaborative vision-text optimizing mechanism within the OVS field. Extensive experiments demonstrate our method achieves superior performance on popular OVS benchmarks. In open-vocabulary semantic segmentation, our method outperforms the previous state-of-the-art approaches by +0.5, +2.3, +3.4, +0.4 and +1.1 mIoU, respectively on A-847, A-150, PC-459, PC-59 and PAS-20. Furthermore, in a panoptic setting on the ADE20K dataset, …

The Segment Anything Model (SAM) has demonstrated remarkable zero-shot capability and flexible geometric prompting in general image segmentation. However, it often struggles in domains that are either sparsely represented or lie outside its training distribution, such as aerial, medical, and non-RGB images. Recent efforts have predominantly focused on adapting SAM to these domains using fully supervised methods, which necessitate large amounts of annotated training data and pose practical challenges in data collection. This paper presents CAT-SAM, a ConditionAl Tuning network that explores few-shot adaptation of SAM toward various challenging downstream domains in a data-efficient manner. The core design is a prompt bridge structure that enables decoder-conditioned joint tuning of the heavyweight image encoder and the lightweight mask decoder. The bridging maps the domain-specific features of the mask decoder to the image encoder, fostering synergic adaptation of both components with mutual benefits with few-shot target samples only, ultimately leading to superior segmentation in various downstream tasks. We develop two CAT-SAM variants that adopt two tuning strategies for the image encoder: one injecting learnable prompt tokens in the input space and the other inserting lightweight adapter networks. Extensive experiments over 11 downstream tasks show that CAT-SAM achieves superior segmentation consistently even under …

Unsupervised Domain Adaptation (UDA) for semantic segmentation has been widely studied to exploit the label-rich source data to assist the segmentation of unlabeled samples on target domain. Despite these efforts, UDA performance remains far below that of fully-supervised model owing to the lack of target annotations. To this end, we propose an efficient superpixel-level active learning method for domain adaptive semantic segmentation to maximize segmentation performance by automatically querying a small number of superpixels for labeling. To conserve annotation resources, we propose a novel low-uncertainty superpixel fusion module which amalgamates superpixels possessing low-uncertainty features based on feature affinity and thereby ensuring high-quality fusion of superpixels. As for the acquisition strategy, our method takes into account two types of information-rich superpixels: large-size superpixels with substantial information content, and superpixels with the greatest value for domain adaptation learning. Further, we employ the cross-domain mixing and pseudo label with consistency regularization techniques respectively to address the domain shift and label noise problems. Extensive experimentation demonstrates that our proposed superpixel-level method utilizes a limited budget more efficiently than previous pixel-level techniques and surpasses state-of-the-art methods at 40x lower cost.

Delving into the realm of egocentric vision, the advancement of referring video object segmentation (RVOS) stands as pivotal in understanding human activities. However, existing RVOS task primarily relies on static attributes such as object names to segment target objects, posing challenges in distinguishing target objects from background objects and in identifying objects undergoing state changes. To address these problems, this work proposes a novel action-aware RVOS setting called ActionVOS, aiming at segmenting only active objects in egocentric videos using human actions as a key language prompt. This is because human actions precisely describe the behavior of humans, thereby helping to identify the objects truly involved in the interaction and to understand possible state changes. We also build a method tailored to work under this specific setting. Specifically, we develop an action-aware labeling module with an efficient action-guided focal loss. Such designs enable ActionVOS model to prioritize active objects with existing readily-available annotations. Experimental results on VISOR dataset reveal that ActionVOS significantly reduces the mis-segmentation of inactive objects, confirming that actions help the ActionVOS model understand objects' involvement. Further evaluations on VOST and VSCOS datasets show that the novel ActionVOS setting enhances segmentation performance when encountering challenging circumstances involving object state …

Segmenting and recognizing a diverse range of object parts is crucial in various computer vision and robotic applications. While object segmentation has made significant progress, part-level segmentation remains an under-explored issue. Part segmentation entails discerning complex boundaries between parts, and the scarcity of annotated data further complicates the task. To tackle this problem, in this paper, we propose a novel Weakly-supervised Part Segmentation (WPS) setting and an approach called WPS-SAM, built on the large-scale pre-trained vision foundation model, Segment Anything Model (SAM). WPS-SAM is an end-to-end framework designed to extract prompt tokens directly from images and perform pixel-level segmentation of part regions. During its training phase, it only utilizes weakly supervised labels in the form of bounding boxes or points. Extensive experiments demonstrate that, through exploiting the rich knowledge embedded in pre-trained foundation models, WPS-SAM outperforms other segmentation models trained with pixel-level strong annotations. Specifically, WPS-SAM achieves 68.93% mIOU and 79.53% mACC on the PartImageNet dataset, surpassing state-of-the-art fully supervised methods by approximately 4% in terms of mIOU.

Robustness is the most important property of watermarking schemes. In practice, the watermarking mechanism shall be robust to both geometric and non-geometric distortions. In deep learning-based watermarking frameworks, robustness can be ensured by end-to-end training with different noise layers. However, most of the current CNN-based watermarking frameworks, even trained with targeted distortions, cannot well adapt to geometric distortions due to the architectural design. Since the traditional convolutional layer's position structure is relatively fixed, it lacks the flexibility to capture the influence of geometric distortion, making it difficult to train for corresponding robustness. To address such limitations, we propose a Swin Transformer and Deformable Convolutional Network (DCN)-based watermark model backbone. The attention mechanism and the deformable convolutional window effectively improve the feature processing flexibility, greatly enhancing the robustness, especially for geometric distortions. Besides, for non-geometric distortions, aiming at improving the generalizability for more distortions, we also provide a distortion-style-ensembled noise layer, including an image encoder, an image decoder, and distortion-style layers that can effectively simulate styles of different kinds of distortions. In the final watermark model training stage, we can simply train with our proposed noise layer for overall robustness. Extensive experiments illustrate that compared to existing state-of-the-art (SOTA) works, with …

The generation of large-scale urban layouts has garnered substantial interest across various disciplines. Prior methods have utilized procedural generation requiring manual rule coding or deep learning needing abundant data. However, prior approaches have not considered the context-sensitive nature of urban layout generation. Our approach addresses this gap by leveraging a canonical graph representation for the entire city, which facilitates scalability and captures the multi-layer semantics inherent in urban layouts. We introduce a novel graph-based masked autoencoder (GMAE) for city-scale urban layout generation. The method encodes attributed buildings, city blocks, communities and cities into a unified graph structure, enabling self-supervised masked training for graph autoencoder. Additionally, we employ scheduled iterative sampling for 2.5D layout generation, prioritizing the generation of important city blocks and buildings. Our approach achieves good realism, semantic consistency, and correctness across the heterogeneous urban styles in 330 US cities. Codes and datasets are released at: https://github.com/Arking1995/COHO.

6-DoF grasp detection has been a fundamental and challenging problem in robotic vision. While previous works have focused on ensuring grasp stability, they often do not consider human intention conveyed through natural language, hindering effective collaboration between robots and users in complex 3D environments. In this paper, we present a new approach for language-driven 6-DoF grasp detection in cluttered point clouds. We first introduce Grasp-Anything-6D, a large-scale dataset for the language-driven 6-DoF grasp detection task with 1M point cloud scenes and more than 200M language-associated 3D grasp poses. We further introduce a novel diffusion model that incorporates a new negative prompt guidance learning strategy. The proposed negative prompt strategy directs the detection process toward the desired object while steering away from unwanted ones given the language input. Our method enables an end-to-end framework where humans can command the robot to grasp desired objects in a cluttered scene using natural language. Intensive experimental results show the effectiveness of our method in both benchmarking experiments and real-world scenarios, surpassing other baselines. In addition, we demonstrate the practicality of our approach in real-world robotic applications.

Predicting realistic ground views from satellite imagery in urban scenes is a challenging task due to the significant view gaps between satellite and ground-view images. We propose a novel pipeline to tackle this challenge, by generating geospecifc views that maximally respect the weak geometry and texture from multi-view satellite images. Different from existing approaches that hallucinate images from cues such as partial semantics or geometry from overhead satellite images, our method directly predicts ground-view images at geolocation by using a comprehensive set of information from the satellite image, resulting in ground-level images with a resolution boost at a factor of ten or more. We leverage a novel building refinement method to reduce geometric distortions in satellite data at ground level, which ensures the creation of accurate conditions for view synthesis using diffusion networks. Moreover, we proposed a novel geospecific prior, which prompts distribution learning of diffusion models to respect image samples that are closer to the geolocation of the predicted images. We demonstrate our pipeline is the first to generate close-to-real and geospecific ground views merely based on satellite images. Codes and data will be shared.

We aim to discover manipulation concepts embedded in the unannotated demonstrations, which are recognized as key physical states. The discovered concepts can facilitate training manipulation policies and promote generalization. Current methods relying on multimodal foundation models for deriving key states usually lack accuracy and semantic consistency due to limited multimodal robot data. In contrast, we introduce an information-theoretic criterion to characterize the regularities that signify a set of physical states. We also develop a framework that trains a concept discovery network using this criterion, thus bypassing the dependence on human semantics and alleviating costly human labeling. The proposed criterion is based on the observation that key states, which deserve to be conceptualized, often admit more physical constraints than non-key states. This phenomenon can be formalized as maximizing the mutual information between the putative key state and its preceding state, i.e., Maximal Mutual Information (MaxMI). By employing MaxMI, the trained key state localization network can accurately identify states of sufficient physical significance, exhibiting reasonable semantic compatibility with human perception. Furthermore, the proposed framework produces key states that lead to concept-guided manipulation policies with higher success rates and better generalization in various robotic tasks compared to the baselines, verifying the effectiveness of …

With the comprehensive research conducted on various face analysis tasks, there is a growing interest among researchers to develop a unified approach to face perception. Existing methods mainly discuss unified representation and training, which lack task extensibility and application efficiency. To tackle this issue, we focus on the unified model structure, exploring a face generalist model. As an intuitive design, Naive Faceptor enables tasks with the same output shape and granularity to share the structural design of the standardized output head, achieving improved task extensibility. Furthermore, Faceptor is proposed to adopt a well-designed single-encoder dual-decoder architecture, allowing task-specific queries to represent new-coming semantics. This design enhances the unification of model structure while improving application efficiency in terms of storage overhead. Additionally, we introduce Layer-Attention into Faceptor, enabling the model to adaptively select features from optimal layers to perform the desired tasks. Through joint training on 13 face perception datasets, Faceptor achieves exceptional performance in facial landmark localization, face parsing, age estimation, expression recognition, binary attribute classification, and face recognition, achieving or surpassing specialized methods in most tasks. Our training framework can also be applied to auxiliary supervised learning, significantly improving performance in data-sparse tasks such as age estimation and expression …

Recent advancements in transformer-based light-weight object tracking have set new standards across various benchmarks due to their efficiency and effectiveness. Despite these achievements, most current trackers rely heavily on pre-existing object detection architectures without optimizing the backbone network to leverage the unique demands of object tracking. Addressing this gap, we introduce the Feature Extraction and Relation Modeling Tracker (FERMT) - a novel approach that significantly enhances tracking speed and accuracy. At the heart of FERMT is a strategic decomposition of the conventional attention mechanism into four distinct sub-modules within a one-stream tracker. This design stems from our insight that the initial layers of a tracking network should prioritize feature extraction, whereas the deeper layers should focus on relation modeling between objects. Consequently, we propose an innovative, light-weight backbone specifically tailored for object tracking. Our approach is validated through meticulous ablation studies, confirming the effectiveness of our architectural decisions. Furthermore, FERMT incorporates a Dual Attention Unit for feature pre-processing, which facilitates global feature interaction across channels and enriches feature representation with attention cues. Benchmarking on GOT-10k, FERMT achieves a groundbreaking Average Overlap (AO) score of 69.6%, outperforming the leading real-time trackers by 5.6% in accuracy while boasting a 54% improvement in …

Multimodal models have gained increasing popularity recently. Many works have been proposed to learn the representations for different modalities. The representation can learn shared information from these domains, leading to increased and coherent joint and cross-generation. However, these works mainly considered standard Gaussian or Laplacian as their prior distribution. It can be challenging for the uni-modal and non-informative distribution to capture all the information from multiple data types. Meanwhile, energy-based models (EBM) have shown their effectiveness in multiple tasks due to their expressiveness and flexibility. But its capacity has yet to be discovered for the multimodal generative models. In this paper, we propose a novel framework to train multimodal latent generative models together with the energy-based models. The proposed method can lead to more expressive and informative prior which can better capture the information within multiple modalities. Our experiments showed that our model is effective and can increase generation coherence and latent classification for different multimodal datasets.

Self-supervised learning (SSL) has proven effective in learning high-quality representations for various downstream tasks, with a primary focus on semantic tasks. However, its application in geometric tasks remains underexplored, partially due to the absence of a standardized evaluation method for geometric representations. To address this gap, we introduce a novel pose-estimation benchmark for assessing SSL geometric representations, which demands training without semantic or pose labels and achieving proficiency in both semantic and geometric downstream tasks. On this benchmark, we study enhancing SSL geometric representations without sacrificing semantic classification accuracy. We find that leveraging mid-layer representations improves pose-estimation performance by 10-20%. Further, we introduce an unsupervised trajectory-regularization loss, which improves performance by an additional 4% and improves generalization ability on out-of-distribution data. We hope the proposed benchmark and methods offer new insights and improvements in self-supervised geometric representation learning.

Vision Transformer models trained on large-scale datasets, although effective, often exhibit artifacts in the patch token they extract. While such defects can be alleviated by re-training the entire model with additional classification tokens, the underlying reasons for the presence of these tokens remain unclear. In this paper, we conduct a thorough investigation of this phenomenon, combining theoretical analysis with empirical observations. Our findings reveal that these artifacts originate from the pre-trained network itself, specifically stemming from the leading left singular vector of the network's weights. Furthermore, to mitigate these defects, we propose a novel fine-tuning smooth regularization that rectifies structural deficiencies using only a small dataset, thereby avoiding the need for complete re-training. We validate our method on various downstream tasks, including unsupervised segmentation, classification, and supervised segmentation, demonstrating its effectiveness in improving model performance. Our code and checkpoints will be released.

While recent vision-and-language models (VLMs) like CLIP are a powerful tool for analyzing text and images in a shared semantic space, they do not explicitly model the hierarchical nature of the set of texts which may describe an image. Conversely, existing multimodal hierarchical representation learning methods require costly training from scratch, failing to leverage the knowledge encoded by state-of-the-art multimodal foundation models. In this work, we study the knowledge of existing foundation models, finding that they exhibit emergent understanding of visual-semantic hierarchies despite not being directly trained for this purpose. We propose the Radial Embedding (RE) framework for probing and optimizing hierarchical understanding, and contribute the HierarCaps dataset, a benchmark facilitating the study of hierarchical knowledge in image--text representations, constructed automatically via large language models. Our results show that foundation VLMs exhibit zero-shot hierarchical understanding, surpassing the performance of prior models explicitly designed for this purpose. Furthermore, we show that foundation models may be better aligned to hierarchical reasoning via a text-only fine-tuning phase, while retaining pretraining knowledge. We will release our data, code, and trained models.

Fueled by the Large Language Models (LLMs) wave, Large Visual-Language Models (LVLMs) have emerged as a pivotal advancement, bridging the gap between image and text. However, video making it challenging for LVLMs to perform adequately due to the complexity of the relationship between language and spatial-temporal data structure. Recent Large Video-Language Models (LVidLMs) align feature of static visual data like image into latent space of language feature, by general multi-modal tasks to leverage abilities of LLMs sufficiently. In this paper, we explore fine-grained alignment approach via object trajectory for different modalities across both spatial and temporal dimensions simultaneously. Thus, we propose a novel LVidLM by trajectory-guided Pixel-Temporal Alignment, dubbed PiTe, that exhibits promising applicable model property. To achieve fine-grained video-language alignment, we curate a multi-modal pre-training dataset PiTe-143k, the dataset provision of moving trajectories in pixel level for all individual objects, that appear and mention in the video and caption both, by our automatic annotation pipeline. Meanwhile, PiTe demonstrates astounding capabilities on myriad video-related multi-modal tasks through beat the state-of-the-art methods by a large margin.

The increasing availability of multi-sensor data sparks interest in multimodal self-supervised learning. However, most existing approaches learn only common representations across modalities while ignoring intra-modal training and modality-unique representations. We propose Decoupling Common and Unique Representations (DeCUR), a simple yet effective method for multimodal self-supervised learning. By distinguishing inter- and intra-modal embeddings through multimodal redundancy reduction, DeCUR can integrate complementary information across different modalities. Meanwhile, a simple residual deformable attention is introduced to help the model focus on modality-informative features. We evaluate DeCUR in three common multimodal scenarios ( radar-optical, RGB-elevation, and RGB-depth), and demonstrate its consistent and significant improvement for both multimodal and modality-missing settings. With thorough experiments and comprehensive analysis, we hope this work can provide insights and raise more interest in researching the hidden relationships of multimodal representations.

We delve into a crucial yet often overlooked challenge inherent to Vision Transformers (ViTs): feature maps of these models exhibit grid-like artifacts, which hurt the performance of ViTs in downstream dense prediction tasks such as segmentation, depth prediction, and object discovery. We trace this fundamental issue down to the positional embeddings at the input stage. we propose a two-stage denoising approach, termed Denoising Vision Transformers (DVT). In the first stage, we separate the clean features from those contaminated by positional artifacts by enforcing cross-view feature consistency with neural fields on a per-image basis. This per-image optimization process extracts artifact-free features from raw ViT outputs, providing clean feature estimates for offline applications. In the second stage, we train a lightweight Transformer block to predict clean features from raw ViT outputs, leveraging the derived estimates of the clean features as supervision. Our DVT does not require re-training the existing pre-trained ViTs, and is immediately applicable to any Vision Transformer architecture. We evaluate our method on a variety of representative ViTs (DINO, DeiT-III, EVA02, CLIP, DINOv2, DINOv2-reg) and demonstrate that our DVT consistently and significantly improves existing state-of-the-art general-purpose models in semantic and geometric tasks across multiple datasets. We hope our study will …

Talking face generation aims to create a realistic video with accurate lip synchronization and high visual quality, using given audio and reference video, while preserving identity and visual characteristics. In this paper, we start by identifying several issues of existing synchronization learning methods. These involve unstable training, lip synchronization, and visual quality issues caused by lip-sync loss and SyncNet. We further tackle lip leaking problem from the identity reference and propose a silent-lip generator, aiming to prevent lip leaking by changing the lips of the identity reference. We then introduce stabilized synchronization loss and AVSyncNet to alleviate the problems caused by lip-sync loss and SyncNet. Finally, we present adaptive triplet loss to enhance visual quality and apply a post-processing technique to obtain high-quality videos. According to the experiments, our model outperforms state-of-the-art methods in both visual quality and lip synchronization. Comprehensive ablation studies further validate our individual contributions as well as their complementary effects.

Speech-driven 3D talking heads generation has emerged as a significant area of interest among researchers, presenting numerous challenges. Existing methods are constrained by animating faces with fixed topologies, wherein point-wise correspondence is established, and the number and order of points remains consistent across all identities the model can animate. In this work, we present ScanTalk, a novel framework capable of animating 3D faces in arbitrary topologies including scanned data. Our approach relies on the DiffusionNet architecture to overcome the fixed topology constraint, offering promising avenues for more flexible and realistic 3D animations. By leveraging the power of DiffusionNet, ScanTalk not only adapts to diverse facial structures but also maintains fidelity when dealing with scanned data, thereby enhancing the authenticity and versatility of generated 3D talking heads. Through comprehensive comparisons with state-of-the-art methods, we validate the efficacy of our approach, demonstrating its capacity to generate realistic talking heads comparable to existing techniques. While our primary objective is to develop a generic method free from topological constraints, all state-of-the-art methodologies are bound by such limitations. Code for reproducing our results, and the pre-trained model will be made available.

In this paper, we propose a novel learning approach for feed-forward one-shot 4D head avatar synthesis. Different from existing methods that often learn from reconstructing monocular videos guided by 3DMM, we employ pseudo multi-view videos to learn a 4D head synthesizer in a data-driven manner, avoiding reliance on inaccurate 3DMM reconstruction that could be detrimental to the synthesis performance. The key idea is to first learn a 3D head synthesizer using synthetic multi-view images to convert monocular real videos into multi-view ones, and then utilize the pseudo multi-view videos to learn a 4D head synthesizer via cross-view self-reenactment. By leveraging a simple vision transformer backbone with motion-aware cross-attentions, our method exhibits superior performance compared to previous methods in terms of reconstruction fidelity, geometry consistency, and motion control accuracy. We hope our method offers novel insights into integrating 3D priors with 2D supervisions for improved 4D head avatar creation.

Virtual Reality (VR) bares promise of social interactions that can feel more immersive than other media. Key to this is the ability to accurately animate a personalized photorealistic avatar, and hence the acquisition of the labels for headset-mounted camera (HMC) images need to be efficient and accurate, while wearing a VR headset. This is challenging due to oblique camera views and differences in image modality. In this work, we first show that the domain gap between the avatar and HMC images is one of the primary sources of difficulty, where a transformer-based architecture achieves high accuracy on domain-consistent data, but degrades when the domain-gap is re-introduced. Building on this finding, we propose a system split into two parts: an iterative refinement module that takes in-domain inputs, and a generic avatar-guided image-to-image domain transfer module conditioned on current estimates. These two modules reinforce each other: domain transfer becomes easier when close-to-groundtruth examples are shown, and better domain-gap removal in turn improves the registration. Our system obviates the need for costly offline optimization, and produces online registration of higher quality than direct regression method. We validate the accuracy and efficiency of our approach through extensive experiments on a commodity headset, demonstrating significant …

We present a novel pipeline for learning triangular human avatars from multi-view videos. Recent methods for avatar learning are typically based on neural radiance fields (NeRF), which is not compatible with traditional graphics pipeline and poses great challenges for operations like editing or synthesizing under different environments. To overcome these limitations, our method represents the avatar with an explicit triangular mesh extracted from an implicit SDF field, complemented by an implicit material field conditioned on given poses. Leveraging this triangular avatar representation, we incorporate physics-based rendering to accurately decompose geometry and material. To enhance both the geometric and appearance details, we further employ a 2D UNet as the network backbone and introduce pseudo normal ground-truth as additional supervision. Experiments show that our method can learn triangular avatars with high-quality geometry reconstruction and material decomposition, inherently supporting editing, manipulation or relighting operations.

In this paper, we present Export3D, a one-shot 3D-aware portrait animation method that is able to control the facial expression and camera view of a given portrait image. To achieve this, we introduce a tri-plane generator that directly generates a tri-plane of 3D prior by transferring the expression parameter of 3DMM into the source image. The tri-plane is then decoded into the image of different view through a differentiable volume rendering. Existing portrait animation methods heavily rely on image warping to transfer the expression in the motion space, challenging on disentanglement of appearance and expression. In contrast, we propose a contrastive pre-training framework for appearance-free expression parameter, eliminating undesirable appearance swap when transferring a cross-identity expression. Extensive experiments show that our pre-training framework can learn the appearance-free expression representation hidden in 3DMM, and our model can generate 3D-aware expression controllable portrait image without appearance swap in the cross-identity manner.

Spike camera with high temporal resolution can fire continuous binary spike streams to record per-pixel light intensity. By using reconstruction methods, the scene details in high-speed scenes can be restored from spike streams. However, existing methods struggle to perform well in low-light environments due to insufficient information in spike streams. To this end, we propose a bidirectional recurrent-based reconstruction framework to better handle such extreme conditions. In more detail, a \textbf{l}ight-\textbf{r}obust \textbf{rep}resentation (LR-Rep) is designed to aggregate temporal information in spike streams. Moreover, a fusion module is used to extract temporal features. Besides, we synthesize a reconstruction dataset for high-speed low-light scenes where light sources are carefully designed to be consistent with reality. The experiment shows the superiority of our method. Importantly, our method also generalizes well to real spike streams. Our project is: https://github.com/Acnext/Learning-to-Robustly-Reconstruct-Dynamic-Scenes-from-Low-light-Spike-Streams/.

Spectral imaging offers the capability to unveil hidden details within the world around us. However, to fully harness this potential, it is imperative to develop effective spectral demosaicing techniques. Despite the success of learning based spectral demosaicing methods, three challenges hinder their practical use. Firstly, existing convolutional neural networks and attention-based models, struggle to capture spectral similarities and long-range dependencies. Secondly, their performance is unstable when optical characteristics, like multispectral filter array (MSFA) arrangement and wavelength distribution, change. Lastly, they lack a structured approach to incorporating imaging system physics, such as MSFA pattern. Addressing these challenges, our paper introduces the Wavelength Embedding guided Filter Array Attention Transformer (WeFAT) for effective spectral demosaicing. Specifically, akin to timestep embedding in denoising diffusion models, we propose a Wavelength Embedding guided Multi-head Self-Attention (We-MSA) mechanism to imbue our model with wavelength memory, facilitating adaptation to diverse cameras. This approach treats each spectral feature as a token, directly integrating wavelength information into attention calculation. Additionally, we developed a MSFA-attention Mechanism (MaM) steering We-MSA to focus on spatial regions yielding high-quality spectral data. Experimental results affirm that WeFAT exhibits strong performance consistency across diverse cameras characterized by varying spectral distributions and MSFA patterns, trained solely on …

High dynamic range imaging records natural scenes with closer to actual luminance distributions, at the cost of increased storage requirements and display demands. Consequently, HDR image compression for perceptually optimal storage and display is crucial, yet it remains inadequately addressed. In this work, we take steps towards this goal. Specifically, we learn to compress HDR images into two bitstreams for storage, one of which is used to generate low dynamic range (LDR) images for display purposes conditioned on the maximum luminance of the scene, while the other serves as side information to aid HDR image reconstruction from the generated LDR image. To measure the perceptual quality of the displayable LDR image, we employ the normalized Laplacian pyramid distance (NLPD), a perceptually quality metric that supports the use of the input HDR image as reference. To measure the perceptual quality of the reconstructed HDR image, we employ a newly proposed HDR quality metric based on a simple inverse display model that enables high-fidelity dynamic range expansion at all luminance levels. Comprehensive qualitative and quantitative comparisons on various HDR scenes demonstrate the perceptual optimality of our learned HDR image compression system for both displayable LDR images and reconstructed HDR images at all …

This paper aims to facilitate more practical NLOS imaging by reducing the number of samplings and scan areas. To this end, we introduce a phasor-based enhancement network that is capable of predicting clean and full measurements from noisy partial observations. We leverage a denoising autoencoder scheme to acquire rich and noise-robust representations in the measurement space. Through this pipeline, our enhancement network is trained to accurately reconstruct complete measurements from their corrupted and partial counterparts. However, we observe that the \naive application of denoising often yields degraded and over-smoothed results, caused by unnecessary and spurious frequency signals present in measurements. To address this issue, we introduce a phasor-based pipeline designed to limit the spectrum of our network to the frequency range of interests, where the majority of informative signals are detected. The phasor wavefronts at the aperture, which are band-limited signals, are employed as inputs and outputs of the network, guiding our network to learn from the frequency range of interests and discard unnecessary information. The experimental results in more practical acquisition scenarios demonstrate that we can look around the corners with 16× or 64× fewer samplings and 4× smaller apertures.

Neural Radiance Fields (NeRF) accomplishes photo-realistic novel view synthesis by learning the implicit volumetric representation of a scene from multi-view images, which faithfully convey the colorimetric information. However, sensor noises will contaminate low-value pixel signals, and the lossy camera image signal processor will further remove near-zero intensities in extremely dark situations, deteriorating the synthesis performance. Existing approaches reconstruct low-light scenes from raw images but struggle to recover texture and boundary details in dark regions. Additionally, they are unsuitable for high-speed models relying on explicit representations. To address these issues, we present Thermal-NeRF, which takes thermal and visible raw images as inputs, considering the thermal camera is robust to the illumination variation and raw images preserve any possible clues in the dark, to accomplish visible and thermal view synthesis simultaneously. Also, the first multi-view thermal and visible dataset (MVTV) is established to support the research on multimodal NeRF. Thermal-NeRF achieves the best trade-off between detail preservation and noise smoothing and provides better synthesis performance than previous work. Finally, we demonstrate that both modalities are beneficial to each other in 3D reconstruction.

Inverse rendering of outdoor scenes from unconstrained image collections is a challenging task, particularly illumination/albedo ambiguities and occlusion of the illumination environment (shadowing) caused by geometry. However, there are many cues in an image that can aid in the disentanglement of geometry, albedo and shadows. Whilst sky is frequently masked out in state-of-the-art methods, we exploit the fact that any sky pixel provides a direct observation of distant lighting in the corresponding direction and, via a neural illumination prior, a statistical cue to derive the remaining illumination environment. The incorporation of our illumination prior is enabled by a novel `outside-in' method for computing differentiable sky visibility based on a neural directional distance function. This is highly efficient and can be trained in parallel with the neural scene representation, allowing gradients from appearance loss to flow from shadows to influence the estimation of illumination and geometry. Our method estimates high-quality albedo, geometry, illumination and sky visibility, achieving state-of-the-art results on the NeRF-OSR relighting benchmark.

Several variants of Neural Radiance Fields (NeRFs) have significantly improved the accuracy of synthesized images and surface reconstruction of 3D scenes/objects. In all of these methods, a key characteristic is that none can train the neural network with every possible input data, specifically, every pixel and potential 3D point along the projection rays due to scalability issues. While vanilla NeRFs uniformly sample both the image pixels and 3D points along the projection rays, some variants focus only on guiding the sampling of the 3D points along the projection rays. In this paper, we leverage the implicit surface representation of the foreground scene and model a probability density function in a 3D image projection space to achieve a more targeted sampling of the rays toward regions of interest, resulting in improved rendering. Additionally, a new surface reconstruction loss is proposed for improved performance. This new loss fully explores the proposed 3D image projection space model and incorporates near-to-surface and empty space components. By integrating our novel sampling strategy and novel loss into any current state-of-the-art neural implicit surface renderer, we achieve more accurate and detailed 3D reconstructions and improved image rendering, especially for the regions of interest in any given scene. …

This work tackles the challenging task of achieving real-time novel view synthesis on various scenes, including highly reflective objects and unbounded outdoor scenes. Existing real-time rendering methods, especially those based on meshes, often have subpar performance in modeling surfaces with rich view-dependent appearances. Our key idea lies in leveraging meshes for rendering acceleration while incorporating a novel approach to parameterize view-dependent information. We decompose the color into diffuse and specular, and model the specular color in the reflected direction based on a neural environment map. Our experiments demonstrate that our method achieves comparable reconstruction quality for highly reflective surfaces compared to state-of-the-art offline methods, while also efficiently enabling real-time rendering on edge devices such as smartphones.

Dynamic Neural Radiance Field (NeRF) from monocular videos has recently been explored for space-time novel view synthesis and achieved excellent results. However, defocus blur caused by depth variation often occurs in video capture, compromising the quality of dynamic reconstruction because the lack of sharp details interferes with modeling temporal consistency between input views. To tackle this issue, we propose \ourmethod, the first dynamic NeRF method designed to restore sharp novel views from defocused monocular videos. We introduce layered Depth-of-Field (DoF) volume rendering to model the defocus blur and reconstruct a sharp NeRF supervised by defocused views. The blur model is inspired by the connection between DoF rendering and volume rendering. The opacity in volume rendering aligns with the inter-layer visibility in layered DoF rendering. To execute the blurring, we modify the layered blur kernel to the ray-based kernel and employ an optimized sparse kernel to gather the input rays efficiently and render the optimized rays with our layered DoF volume rendering. We synthesize a dataset with defocused dynamic scenes for our task, and extensive experiments on our dataset show that our method outperforms existing approaches in synthesizing all-in-focus novel views from defocus blur while maintaining spatial-temporal consistency in the scene.

The acquisition of multi-task (MT) labels in 3D scenes is crucial for a wide range of real-world applications. Traditional methods generally employ an analysis-by-synthesis approach, generating 2D label maps on novel synthesized views, or utilize Neural Radiance Field (NeRF), which concurrently represents label maps. Yet, these approaches often struggle to balance inference efficiency with MT label quality. Specifically, they face limitations such as (a) constrained rendering speeds due to NeRF pipelines, and (b) the implicit representation of MT fields that can result in continuity artifacts during rendering. Recently, 3D Gaussian Splatting has shown promise in achieving real-time rendering speeds without compromising rendering quality. In our research, we address the challenge of enabling 3D Gaussian Splatting to represent Versatile MT labels. Simply attaching MT attributes to explicit Gaussians compromises rendering quality due to the lack of cross-task information flow during optimization. We introduce architectural and rasterizer design to effectively overcome this issue. Our VersatileGaussian model innovatively associates Gaussians with shared MT features and incorporates a feature map rasterizer. The cornerstone of this versatile rasterization is the Task Correlation Attention module, which fosters cross-task correlations through a soft weighting mechanism that disseminates task-specific knowledge. Across experiments on the ScanNet and Replica datasets …

Neural Radiance Fields (NeRF) have achieved remarkable progress on dynamic scenes with deformable objects. Nonetheless, most of the previous works required multi-view inputs or long training time (several hours), making it hard to apply them for real-world scenarios. Recently, a series of works have been dedicated to addressing the blurry artifact present in synthesized novel views given a monocular input of dynamic scenes. However, they may fail to predict stable and accurate deformation while keeping high-frequency details when rendering at various resolutions. To this end, we introduce a novel framework DMiT (Deformable Mipmapped Tri-Plane) that adopts the mipmaps to render dynamic scenes with various resolutions from novel views. With the help of hierarchical mipmapped triplanes, we incorporate an MLP to effectively predict a mapping between the observation space and the canonical space, enabling not only high-fidelity dynamic scene rendering but also high-performance training and inference. Moreover, a training scheme for joint geometry and deformation refinement is designed for canonical regularization to reconstruct high-quality geometries. Extensive experiments on both synthetic and real dynamic scenes demonstrate the efficacy and efficiency of our method.

In this paper, we first revisit the existing approach of decomposing large-scale scenes into multiple independently trained Neural Radiance Fields (NeRFs), and identify several fundamental issues that hinder performance improvement with additional computational resources (GPUs), which contradicts the fundamental objective of leveraging multi-GPU setups to enhance large-scale NeRF performance. Subsequently, we introduce NeRF-XL, a principled algorithm designed to efficiently harness multi-GPU setups for performance improvement, thereby enabling NeRF at any scale. At its core, our method allocates non-overlapping NeRFs across disjoint spatial regions and optimizes them jointly across GPUs. We reduce the GPU communication overhead by rewriting the volume rendering equation and relevant loss terms, enhancing training and rendering efficiency. Without any heuristics, our approach gracefully reveals scaling laws for NeRFs in the multi-GPU setting across various types of data and scales, including the first time NeRF reconstruction on the largest open-source dataset to date, MatrixCity, with 258K images covering a 25km^2 city area.

Neural Radiance Field (NeRF) methodologies have garnered considerable interest, particularly with the introduction of grid-based feature encoding (GFE) approaches such as Instant-NGP and TensoRF. Conventional NeRF employs positional encoding (PE) and represents a scene with a Multi-Layer Perceptron (MLP). Frequency regularization has been identified as an effective strategy to overcome primary challenges in PE-based NeRFs, including dependency on known camera poses and the requirement for extensive image datasets. While several studies have endeavored to extend frequency regularization to GFE approaches, there is still a lack of basic theoretical foundations for these methods. Therefore, we first clarify the underlying mechanisms of frequency regularization. Subsequently, we conduct a comprehensive investigation into the expressive capability of GFE-based NeRFs and attempt to connect frequency regularization with GFE methods. Moreover, we propose a generalized strategy, G2fR: Generalized Grid-based Frequency Regularization, to address issues of camera pose optimization and few-shot reconstruction with GFE methods. We validate the efficacy of our methods through an extensive series of experiments employing various representations across diverse scenarios.

3D surface reconstruction from multi-view images is essential for scene understanding and interaction. However, complex indoor scenes pose challenges such as ambiguity due to limited observations. Recent implicit surface representations, such as Neural Radiance Fields (NeRFs) and signed distance functions (SDFs), employ various geometric priors to resolve the lack of observed information. Nevertheless, their performance heavily depends on the quality of the pre-trained geometry estimation models. To ease such dependence, we propose regularizing the geometric modeling by explicitly encouraging the mutual information between surface normals of two highly correlated points. In this way, the geometry learning process is modulated by the second-order correlations from noisy (first-order) geometric priors, thus eliminating the bias due to poor generalization. Additionally, we introduce a simple yet effective scheme that utilizes semantic and geometric features to identify correlated points, enhancing their mutual information accordingly. The proposed technique can serve as a plugin for SDF-based neural surface representations. Our experiments demonstrate the effectiveness of the proposed in improving the surface reconstruction quality of major states of the arts, and we will make our code publicly available to support future research.

3D Gaussian Splatting showcases notable advancements in photo-realistic and real-time novel view synthesis. However, it faces challenges in modeling mirror reflections, which exhibit substantial appearance variations from different viewpoints. To tackle this problem, we present MirrorGaussian, the first method for mirror scene reconstruction with real-time rendering based on 3D Gaussian Splatting. The key insight is grounded on the mirror symmetry between the real-world space and the virtual mirror space. We introduce an intuitive dual-rendering strategy that enables differentiable rasterization of both the real-world 3D Gaussians and the mirrored counterpart obtained by reflecting the former about the mirror plane. All 3D Gaussians are jointly optimized with the mirror plane in an end-to-end framework. MirrorGaussian achieves high-quality and real-time rendering in scenes with mirrors, empowering scene editing like adding new mirrors and objects. Comprehensive experiments on multiple datasets demonstrate that our approach significantly outperforms existing methods, achieving state-of-the-art results.

The utilization of the triplane-based radiance fields has gained attention in recent years due to its ability to effectively disentangle 3D scenes with a high-quality representation and low computation cost. A key requirement of this method is the precise input of camera poses. However, due to the local update property of the triplane, a similar joint estimation as previous joint pose-NeRF optimization works easily results in local minima. To this end, we propose the Disentangled Triplane Generation module to introduce global feature context and smoothness into triplane learning, which mitigates errors caused by local updating. Then, we propose the Disentangled Plane Aggregation to mitigate the entanglement caused by the common triplane feature aggregation during camera pose updating. In addition, we introduce a two-stage warm-start training strategy to reduce the implicit constraints caused by the triplane generator. Quantitative and qualitative results demonstrate that our proposed method achieves state-of-the-art performance in novel view synthesis with noisy or unknown camera poses, as well as efficient convergence of optimization. Code will be available soon.

The field of 3D reconstruction from images has rapidly evolved in the past few years, first with the introduction of Neural Radiance Field (NeRF) and more recently with 3D Gaussian Splatting (3DGS). The latter provides a significant edge over NeRF in terms of fast training and real-time inference while improving the reconstruction quality. Although the current 3DGS approach works well for dense input images, the unstructured point-cloud like representation quickly overfits to the more challenging setup of sparse training images (e.g., 3 images), creating a representation that appears as a jumble of needles from novel views. We propose to solve this issue by regularized optimization and depth-based initialization. Specifically, we optimize the Gaussian blobs to smoothly and independently deform different object surfaces to compensate for the inaccuracies of the initialization by utilizing an implicit convolutional decoder and a total variation loss. To support our regularized optimization, we initialize a 3D Gaussian representation from each input view through a novel technique that utilizes monocular depth. We demonstrate significant improvements in terms of recovering scene geometry and texture compared to state-of-the-art sparse-view NeRF-based approaches on a variety of scenes.

Implicit neural representation methods have shown impressive advancements in learning 3D scenes from unstructured in-the-wild photo collections but are still limited by the large computational cost of volumetric rendering. More recently, 3D Gaussian Splatting emerged as a much faster alternative with superior rendering quality and training efficiency, especially for small-scale and object-centric scenarios. Nevertheless, this technique suffers from poor performance on unstructured in-the-wild data. To tackle this, we extend over 3D Gaussian Splatting to handle unstructured image collections. We achieve this by modeling appearance to seize photometric variations in the rendered images. Additionally, we introduce a new mechanism to train transient Gaussians to handle the presence of scene occluders in an unsupervised manner. Experiments on diverse photo collection scenes and multi-pass acquisition of outdoor landmarks show the effectiveness of our method over prior works achieving state-of-the-art results with improved efficiency.

This paper presents a novel approach for surface mesh reconstruction from 3D Gaussian Splatting (3DGS), a technique renowned for its efficiency in novel view synthesis but challenged for surface reconstruction. The key obstacle is the lack of geometry hints to regulate the optimization of millions of unorganized Gaussian blobs to align to the true surface. This paper introduces local structural hints during training to address the challenge. We first leverage the prior knowledge from monocular normal and depth estimations to refine the covariance and mean of Gaussian primitives, enhancing their organization and providing crucial normal information for surface extraction. However, due to the highly discrete nature of Gaussian primitives, such geometry guidance remains insufficient for the alignment with the true surface. We then propose to construct a signed distance field by a moving least square (MLS) function over the Gaussians in each local region. More importantly, we further propose to jointly learn a neural implicit network to mimic and regularize the MLS function. The joint optimization helps the optimization of Gassuain Splatting towards accurate surface alignment. Extensive experimental results demonstrate the effectiveness of our method in achieving superior mesh quality compared with the SoTA surface reconstruction for 3DGS. Our code …

3D Gaussian Splatting (3DGS) has demonstrated impressive novel view synthesis results while advancing real-time rendering performance. However, it relies heavily on the quality of the initial point cloud, resulting in blurring and needle-like artifacts in areas with insufficient initializing points. This is mainly attributed to the point cloud growth condition in 3DGS that only considers the average gradient magnitude of points from observable views, thereby failing to grow enough points in these areas. To this end, we propose a novel method, named Pixel-GS, to take into account the number of pixels covered by the Gaussian in each view during the computation of the growth condition. We regard the covered pixel numbers as the weights to dynamically average the gradients from different views. Doing so can mitigate the issue of large Gaussians participating in calculations across too many viewpoints, where a significant number of these viewpoints involve calculations for only a few pixels in boundary regions, leading to smaller gradients and consequently lowering the average gradient magnitude. As a result, points within the areas with insufficient initializing points can be grown more effectively, leading to a more accurate and detailed reconstruction. Besides, we also propose a simple yet effective strategy to …

We propose GS-LRM, a scalable large reconstruction model that can predict high-quality 3D Gaussian primitives from 2-4 posed sparse images in ~0.23 seconds on single A100 GPU. Our model features a very simple transformer-based architecture; we patchify input posed images, pass the concatenated multi-view image tokens through a sequence of transformer blocks, and decode final per-pixel Gaussian parameters directly from these tokens for differentiable rendering. In contrast to previous LRMs that can only reconstruct objects, by predicting per-pixel Gaussians, GS-LRM naturally handles scenes with large variations in scale and complexity. We show that our model can work on both object and scene captures by training it on Objaverse and RealEstate10K respectively. In both scenarios, the models outperform state-of-the-art baselines by a wide margin. We also demonstrate applications of our model in downstream 3D generation tasks.

Novel view synthesis has shown rapid progress recently, with methods capable of producing increasingly photorealistic results. 3D Gaussian Splatting has emerged as a promising method, producing high-quality renderings of scenes and enabling interactive viewing at real-time frame rates. However, it is limited to static scenes. In this work, we extend 3D Gaussian Splatting to reconstruct dynamic scenes. We model a scene's dynamics using dynamic MLPs, learning deformations from temporally-local canonical representations to per-frame 3D Gaussians. To disentangle static and dynamic regions, tuneable parameters weigh each Gaussian's respective MLP parameters, improving the dynamics modelling of imbalanced scenes. We introduce a sliding window training strategy that partitions the sequence into smaller manageable windows to handle arbitrary length scenes while maintaining high rendering quality. We propose an adaptive sampling strategy to determine appropriate window size hyperparameters based on the scene's motion, balancing training overhead with visual quality. Training a separate dynamic 3D Gaussian model for each sliding window allows the canonical representation to change, enabling the reconstruction of scenes with significant geometric changes. Temporal consistency is enforced using a fine-tuning step with self-supervising consistency loss on randomly sampled novel views. As a result, our method produces high-quality renderings of general dynamic scenes with …

Reconstructing surfaces from normals is a key component of photometric stereo. This work introduces an adaptive surface triangulation in the image domain and performs the normal integration on a triangular mesh afterwards. Our key insight is that surface curvature can be computed from normals. Based on curvature, we identify flat areas and aggregate pixels into triangles. The approximation quality is controlled by a single user parameter facilitating a seamless generation of low to high-resolution meshes. Compared to pixel grids, our triangle meshes adapt locally to surface details and allow for a sparser representation. Our new mesh-based formulation of the normal integration problem is strictly derived from discrete differential geometry and leads to well-conditioned linear systems. Results on real and synthetic data show that 10 to 100 times fewer vertices than pixels can be achieved. Experiments suggest that this sparsity translates into a sublinear runtime in the number of pixels. For 64 MP normal maps, our meshing first approach generates and integrates meshes in minutes while pixel-based approaches require hours for the integration alone.

Novel view synthesis from an in-the-wild video is difficult due to challenges like scene dynamics and lack of parallax. While existing methods have shown promising results with implicit neural radiance fields, they are slow to train and render. This paper revisits explicit video representations to synthesize high-quality novel views from a monocular video efficiently. We treat static and dynamic video content separately. Specifically, we build a global static scene model using an extended plane-based scene representation to synthesize temporally coherent novel video. Our plane-based scene representation is augmented with spherical harmonics and displacement maps to capture view-dependent effects and model non-planar complex surface geometry. We opt to represent the dynamic content as per-frame point clouds for efficiency. While such representations are inconsistency-prone, minor temporal inconsistencies are perceptually masked due to motion. We develop a method to quickly estimate such a hybrid video representation and render novel views in real time. Our experiments show that our method can render high-quality novel views from an in-the-wild video with comparable quality to state-of-the-art methods while being 100x faster in training and enabling real-time rendering.

We present 4Diff, a 3D-aware diffusion model addressing the exo-to-ego viewpoint translation problem. This task involves generating first-person (egocentric) view images from third-person (exocentric) images. Leveraging the diffusion model's ability to generate photorealistic images, we propose a transformer-based diffusion model that incorporates geometry priors via the proposed mechanisms: (i) egocentric prior rendering and (ii) 3D-aware rotary cross-attention. The former integrates egocentric layout cues through point cloud rasterization, while the latter incorporates exocentric semantic features by guiding attention between diffusion model feature maps and exocentric semantic features, considering their geometric relationships. Our experiments on the challenging and diverse Ego-Exo4D multiview dataset demonstrate superior performance compared to state-of-the-art approaches. Notably, our approach exhibits robust generalization to novel environments not encountered during training. The code and pretrained models will be made public.

We introduce GeoWizard, a new generative foundation model designed for estimating geometric attributes, e.g., depth and normals, from single images. While significant research has already been conducted in this area, the progress has been substantially limited by the low diversity and poor quality of publicly available datasets. As a result, the prior works either are constrained to limited scenarios or suffer from the inability to capture geometric details. In this paper, we demonstrate that generative models, as opposed to traditional discriminative models (e.g., CNNs and Transformers), can effectively address the inherently ill-posed problem. We further show that leveraging diffusion priors can markedly improve generalization, detail preservation, and efficiency in resource usage. Specifically, we extend the original stable diffusion model to jointly predict depth and normal, allowing mutual information exchange and high consistency between the two representations. More importantly, we propose a simple yet effective strategy to segregate the complex data distribution of various scenes into distinct sub-distributions. This strategy enables our model to recognize different scene layouts, capturing 3D geometry with remarkable fidelity. GeoWizard sets new benchmarks for zero-shot depth and normal prediction, significantly enhancing many downstream applications such as 3D reconstruction, 2D content creation, and novel viewpoint synthesis.

Text-to-image diffusion models generate impressive and realistic images, but do they learn to represent the 3D world from only 2D supervision? We demonstrate that yes, certain 3D scene representations are encoded in the text embedding space of models like Stable Diffusion. Our approach, Viewpoint Neural Textual Inversion (ViewNeTI), is to discover \textit{3D view tokens}; these tokens control the 3D viewpoint --- the rendering pose in a scene --- of generated images. Specifically, we train a small neural mapper to take continuous camera viewpoint parameters and predict a view token (a word embedding); this token conditions diffusion generation via cross-attention to produce images with the desired camera viewpoint. Using ViewNeTI as an evaluation tool, we report two findings: first, the text latent space has a continuous view-control manifold for particular 3D scenes; second, we find evidence for a generalized view-control manifold for all scenes. We conclude that since the view token controls the 3D `rendering' viewpoint, there is likely a scene representation embedded in frozen 2D diffusion models. Finally, we exploit the 3D scene representations for 3D vision tasks, namely, view-controlled text-to-image generation, and novel view synthesis from a single image, where our approach sets state-of-the-art for LPIPS.

Generating high-quality 3D assets from a given image is highly desirable in various applications such as AR/VR. Recent advances in single-image 3D generation explore feed-forward models that learn to infer the 3D model of an object without optimization. Though promising results have been achieved in single object generation, these methods often struggle to model complex 3D assets that inherently contain multiple objects. In this work, we present ComboVerse, a 3D generation framework that produces high-quality 3D assets with complex compositions by learning to combine multiple models. 1) We first perform an in-depth analysis of this "multi-object gap" from both model and data perspectives. 2) Next, with reconstructed 3D models of different objects, we seek to adjust their sizes, rotation angles, and locations to create a 3D asset that matches the given image. 3) To automate this process, we apply spatially-aware score distillation sampling (SSDS) from pretrained diffusion models to guide the positioning of objects. Our proposed framework emphasizes spatial alignment of objects, compared with standard score distillation sampling, and thus achieves more accurate results. Extensive experiments validate ComboVerse achieves clear improvements over existing methods in generating compositional 3D assets.

The field of neural rendering has witnessed significant progress with advancements in generative models and differentiable rendering techniques. Though 2D diffusion has achieved success, a unified 3D diffusion pipeline remains unsettled. This paper introduces a novel framework called LN3Diff to address this gap and enable fast, high-quality, and generic conditional 3D generation. Our approach harnesses a 3D-aware architecture and variational autoencoder (VAE) to encode the input image into a structured, compact, and perceptually equivalent latent space. The latent is decoded by a transformer-based decoder into a high-capacity 3D neural field. Through training a diffusion model on this 3D-aware latent space, our method achieves state-of-the-art performance on ShapeNet for 3D generation and demonstrates superior performance in monocular 3D reconstruction and conditional 3D generation across various datasets. Moreover, it surpasses existing 3D diffusion methods in terms of inference speed, requiring no per-instance optimization. Our proposed LN3Diff presents a significant advancement in 3D generative modeling and holds promise for various applications in 3D vision and graphics tasks.

Generating realistic 3D scenes is challenging due to the complexity of room layouts and object geometries. We propose a sketch based knowledge enhanced diffusion architecture (SEK) for generating customized, diverse, and plausible 3D scenes. SEK conditions the denoising process with a hand-drawn sketch of the target scene and cues from an object relationship knowledge base. We first construct an external knowledge base containing object relationships and then leverage knowledge enhanced graph reasoning to assist our model in understanding hand-drawn sketches. A scene is represented as a combination of 3D objects and their relationships, and then incrementally diffused to reach a Gaussian distribution. We propose a 3D denoising scene transformer that learns to reverse the diffusion process, conditioned by a hand-drawn sketch along with knowledge cues, to regressively generate the scene including the 3D object instances as well as their layout. Experiments on the 3D-FRONT dataset show that our model improves FID, CKL by 17.41\%, 37.18\% in 3D scene generation and FID, KID by 19.12\%, 20.06\% in 3D scene completion compared to the nearest competitor DiffuScene.

We present EchoScene, an interactive and controllable generative model that generates 3D indoor scenes on scene graphs. EchoScene leverages a dual-branch diffusion model that dynamically adapts to scene graphs. Existing methods struggle to handle scene graphs due to varying numbers of nodes, multiple edge combinations, and manipulator-induced node-edge operations. EchoScene overcomes this by associating each node with a denoising process and enables collaborative information exchange, enhancing controllable and consistent generation aware of global constraints. This is achieved through an information echo scheme in both shape and layout branches. At every denoising step, all processes share their denoising data with an information exchange unit that combines these updates using graph convolution. The scheme ensures that the denoising processes are influenced by a holistic understanding of the scene graph, facilitating the generation of globally coherent scenes. The resulting scenes can be manipulated during inference by editing the input scene graph and sampling the noise in the diffusion model. Extensive experiments validate our approach, which maintains scene controllability and surpasses previous methods in generation fidelity. Moreover, the generated scenes are of high quality and thus directly compatible with off-the-shelf texture generation. Code and trained models will be released.

We introduce 3DEgo to address a novel problem of directly synthesizing photorealistic 3D scenes from monocular videos guided by textual prompts. Conventional methods construct a text-conditioned 3D scene through a three-stage process, involving pose estimation using Structure-from-Motion (SfM) libraries like COLMAP, initializing the 3D model with unedited images, and iteratively updating the dataset with edited images to achieve a 3D scene with text fidelity. Our framework streamlines the conventional multi-stage 3D editing process into a single-stage workflow by overcoming the reliance on COLMAP and eliminating the cost of model initialization. We apply a diffusion model to edit video frames prior to 3D scene creation by incorporating our designed \textit{noise blender module} for enhancing multi-view editing consistency, a step that does not require additional training or fine-tuning of T2I diffusion models. 3DEgo utilizes 3D Gaussian Splatting to create 3D scenes from the multi-view consistent edited frames, capitalizing on the inherent temporal continuity and explicit point cloud data. 3DEgo demonstrates remarkable editing precision, speed, and adaptability across a variety of video sources, as validated by extensive evaluations on six datasets, including our own prepared GS25 dataset.

Text-driven 3D texturing requires the generation of high-fidelity texture that conforms to given geometry and description. Recently, the high-quality text-to-image generation ability of 2D diffusion model has significantly promoted this task, by converting it into a texture optimization process guided by multi-view synthesized images, where the generation of high-quality and multi-view consistency images becomes the key issue. State-of-the-art methods achieve the consistency between different views by treating image generation on a novel view as image inpainting conditioned on the texture generated by previously views. However, due to the accumulated semantic divergence of local inpainting and the occlusion between object parts on sparse views, these inpainting-based methods often fail to deal with long-range texture consistency. To address these, we present P3G, a texturing approach based on learned Pseudo 3D Guidance. The key idea of P3G is to first learn a coarse but consistent texture, to serve as a global semantics guidance for encouraging the consistency between images generated on different views. To this end, we incorporate pre-trained text-to-image diffusion models and multi-view optimization to achieve propagating accurate semantics globally for leaning the guidance, and design an efficient framework for high-quality and multi-view consistent image generation that integrates the learned semantic guidance. …

Score Distillation Sampling (SDS) by well-trained 2D diffusion models has shown great promise in text-to-3D generation. However, this paradigm distills view-agnostic 2D image distributions into the rendering distribution of 3D representation for each view independently, overlooking the coherence across views and yielding 3D inconsistency in generations. In this work, we propose \textbf{J}oint \textbf{S}core \textbf{D}istillation (JSD), a new paradigm that ensures coherent 3D generations. Specifically, we model the joint image distribution, which introduces an energy function to capture the coherence among denoised images from the diffusion model. We then derive the joint score distillation on multiple rendered views of the 3D representation, as opposed to a single view in SDS. In addition, we instantiate three universal view-aware models as energy functions, demonstrating compatibility with JSD. Empirically, JSD significantly mitigates the 3D inconsistency problem in SDS, while maintaining text congruence. Moreover, we introduce the Geometry Fading scheme and Classifier-Free Guidance (CFG) Switching strategy to enhance generative details. Our framework, JointDreamer, establishes a new benchmark in text-to-3D generation, achieving outstanding results with an 88.5\% CLIP R-Precision and 27.7\% CLIP Score. These metrics demonstrate exceptional text congruence, as well as remarkable geometric consistency and texture fidelity.

Text-to-3D synthesis has recently emerged as a new approach to sampling 3D models by adopting pretrained text-to-image models as guiding visual priors. An intriguing but underexplored problem with existing text-to-3D methods is that 3D models obtained from the sampling-by-optimization procedure tend to have mode collapses, and hence poor diversity in their results. In this paper, we provide an analysis and identify potential causes of such a limited diversity, which motivates us to devise a new method that considers the joint generation of different 3D models from the same text prompt. We propose to use augmented text prompts via textual inversion of reference images to diversify the joint generation. We show that our method leads to improved diversity in text-to-3D synthesis qualitatively and quantitatively.

Shape abstraction is an important task for simplifying complex geometric structures while retaining essential features. Sweep surfaces, commonly found in human-made objects, aid in this process by effectively capturing and representing object geometry, thereby facilitating abstraction. In this paper, we introduce SweepNet, a novel approach to shape abstraction through sweep surfaces. We propose an effective parameterization for sweep surfaces, utilizing superellipses for profile representation and B-spline curves for the axis. This compact representation, requiring as few as 14 float numbers, facilitates intuitive and interactive editing while preserving shape details effectively. Additionally, by introducing a differentiable neural sweeper and an encoder-decoder architecture, we demonstrate the ability to predict swept volume representations without supervision. We show the superiority of our model through several quantitative and qualitative experiments throughout the paper.

Parametric Computer-Aided Design (CAD) is central to contemporary mechanical design. We harness the capabilities of pre-trained foundation models, renowned for their successes in natural language processing and computer vision, to develop generative models specifically for CAD. These models are adept at understanding complex geometries and design reasoning, a crucial advancement in CAD technology. In this paper, we propose CadVLM, an end-to-end vision language model for CAD generation. Our approach involves adapting pre-trained foundation models to manipulate engineering sketches effectively, integrating both sketch primitive sequences and sketch images. Extensive experiments demonstrate superior performance on multiple CAD sketch generation tasks such as CAD autocompletion, CAD autoconstraint, and image conditional generation. To our knowledge, this is the first instance of a multimodal Large Language Model (LLM) being successfully applied to parametric CAD generation, representing a pioneering step in the field of computer-aided mechanical design. The code is available at https://anonymous.4open.science/r/CadVLM.

This paper introduces a novel framework for virtual try-on, termed Wear-Any-Way. Different from previous methods, Wear-Any-Way is “customizable”. Besides generating high-fidelity results, our method supports users to precisely control the wearing style. To achieve this goal, we first construct a strong pipeline, supporting single/multiple garment try-on and model-to model try-on in complicated scenarios. To make it manipulable, we propose sparse correspondence alignment and involve point-based control to guide the generation. Wear-Any-Way gets state- of-the-art performance for the standard setting and provides a novel interaction form for customizing the wearing style. For instance, it supports users to drag the sleeve to make it rolled up, drag the coat to make it open, and utilize clicks to control the style of tuck, etc. Wear-Any-Way enables more liberated and flexible expressions of the attires, which holds profound implications in the fashion industry.

This paper presents DiffSurf, a transformer-based denoising diffusion model for generating and reconstructing 3D surfaces. Specifically, we design a diffusion transformer architecture that predicts noise from noisy 3D surface vertices and normals. With this architecture, DiffSurf is able to generate 3D surface meshes in various poses and shapes, such as human bodies, hands, animals and man-made objects. Further, DiffSurf is versatile in that it can address various 3D downstream tasks including morphing, body shape variation and 3D human mesh fitting to 2D keypoints. Experimental results on 3D human model benchmarks demonstrate that DiffSurf can generate shapes with greater diversity and higher quality than previous generative models. Furthermore, when applied to the task of single-image 3D human mesh recovery, DiffSurf achieves accuracy comparable to prior techniques at a near real-time rate.

This paper introduces Motion-oriented Compositional Neural Radiance Fields (MoCo-NeRF), a framework designed to perform free-viewpoint rendering of monocular human videos via novel non-rigid motion modeling approach. In the context of dynamic clothed humans, complex cloth dynamics generate non-rigid motions that are intrinsically distinct from skeletal articulations and critically important for the rendering quality. The conventional approach models non-rigid motions as spatial (3D) deviations in addition to skeletal transformations. However, it is either time-consuming or challenging to achieve optimal quality due to its high learning complexity without a direct supervision. To target this problem, we propose a novel approach of modeling non-rigid motions as radiance residual fields to benefit from more direct color supervision in the rendering and utilize the rigid radiance fields as a prior to reduce the complexity of the learning process. Our approach utilizes a single multiresolution hash encoding (MHE) to concurrently learn the canonical T-pose representation from rigid skeletal motions and the radiance residual field for non-rigid motions. Additionally, to further improve both training efficiency and usability, we extend MoCo-NeRF to support simultaneous training of multiple subjects within a single framework, thanks to our effective design for modeling non-rigid motions. This scalability is achieved through the integration …

Neural Radiance Fields (NeRFs) have revolutionized the reconstruction of static scenes and objects in 3D, offering unprecedented quality. However, extending NeRFs to model dynamic objects or object articulations remains a challenging problem. Previous works have tackled this issue by focusing on part-level reconstruction and motion estimation for objects, but they often rely on heuristics regarding the number of moving parts or object categories, which can limit their practical use. In this work, we introduce LEIA, a novel approach for representing dynamic 3D objects. Our method involves observing the object at distinct time steps or "states" and conditioning a hypernetwork on the current state, using this to parameterize our NeRF. This approach allows us to learn a view-invariant latent representation for each state. We further demonstrate that by interpolating between these states, we can generate novel articulation configurations in 3D space that were previously unseen. Our experimental results highlight the effectiveness of our method in articulating objects in a manner that is independent of the viewing angle and joint configuration. Notably, our approach outperforms previous methods that rely on motion information for articulation registration.

We propose a novel neural network approach, LARP (Learned Articulated Rigid body Physics), to model the dynamics of articulated human motion with contact. Our goal is to develop a faster and more convenient methodological alternative to traditional physics simulators for use in computer vision tasks such as human motion reconstruction from video. To that end we introduce a training procedure and model components that support the construction of a recurrent neural architecture to accurately simulate articulated rigid body dynamics. Our neural architecture supports features typically found in traditional physics simulators, such as modeling of joint motors, variable dimensions of body parts, contact between body parts and objects, and is an order of magnitude faster than traditional systems when multiple simulations are run in parallel. To demonstrate the value of LARP we use it as a drop-in replacement for a state of the art classical non-differentiable simulator in an existing video-based reconstruction framework and show comparative or better 3D human pose reconstruction accuracy.

Pre-training video transformers generally requires a large amount of data, presenting significant challenges in terms of data collection costs and concerns related to privacy, licensing, and inherent biases. Synthesizing data is one of the promising ways to solve these issues, yet pre-training solely on synthetic data has its own challenges. In this paper, we introduce an effective self-supervised learning framework for videos that leverages readily available and less costly static images. Specifically, we define the Pseudo Motion Generator (PMG) module that recursively applies image transformations to generate pseudo-motion videos from images. These pseudo-motion videos are then leveraged in masked video modeling. Our approach is applicable to synthetic images as well, thus entirely freeing video pre-training from data collection costs and other concerns in real data. Through experiments in action recognition tasks, we demonstrate that this framework allows effective learning of spatio-temporal features through pseudo-motion videos, significantly improving over existing methods which also use static images and partially outperforming those using both real and synthetic videos. These results uncover fragments of what video transformers learn through masked video modeling.

We embark on the age-old quest: unveiling the hidden dimensions of objects from mere glimpses of their visible parts. To address this,we present \textbf{Vista3D}, a framework that realizes swift and consistent 3D generation within a mere 5 minutes. At the heart of Vista3D lies a two-phase approach: the coarse phase and the fine phase. In the coarse phase, we rapidly generate initial geometry with Gaussian Splatting from a single image. In the fine phase, we extract a Signed Distance Function (SDF) directly from learned Gaussian Splatting, optimizing it with a differentiable isosurface representation. Furthermore, it elevates the quality of generation by using a disentangled representation with two independent implicit functions to capture both visible and obscured aspects of objects. Additionally, it harmonizes gradients from 2D diffusion prior with 3D-aware diffusion priors by angular diffusion prior composition. Through extensive evaluation, we demonstrate that Vista3D effectively sustains a balance between the consistency and diversity of the generated 3D objects. We will make all code and results publicly available.

Establishing reliable correspondences is essential for registration tasks such as 3D and 2D3D registration. Existing methods commonly leverage geometric or semantic point features to generate potential correspondences. However, these features may face challenges such as large deformation, scale inconsistency, and ambiguous matching problems (e.g., symmetry). Additionally, many previous methods, which rely on single-pass prediction, may struggle with local minima in complex scenarios. To mitigate these challenges, we introduce a diffusion matching model for robust correspondence construction. Our approach treats correspondence estimation as a denoising diffusion process within the doubly stochastic matrix space, which gradually denoises (refines) a doubly stochastic matching matrix to the ground-truth one for high-quality correspondence estimation. It involves a forward diffusion process that gradually introduces Gaussian noise into the ground truth matching matrix and a reverse denoising process that iteratively refines the noisy matching matrix. In particular, the feature extraction from the backbone occurs only once during the inference phase. Our lightweight denoising module utilizes the same feature at each reverse sampling step. Evaluation of our method on both 3D and 2D3D registration tasks confirms its effectiveness. The code will be made available online.

Aligning a template to 3D human point clouds is a long-standing problem crucial for tasks like animation, reconstruction, and enabling supervised learning pipelines. Recent data-driven methods leverage predicted surface correspondences; however, they are not robust to varied poses, identities, or noise. In contrast, industrial solutions often rely on expensive manual annotations or multi-view capturing systems. Recently, neural fields have shown promising results. Still, their purely data-driven and extrinsic nature does not incorporate any guidance toward the target surface, often resulting in a trivial misalignment of the template registration. Currently, no method can be considered the standard for 3D Human registration, limiting the scalability of downstream applications. In this work, we propose NSR, a pipeline that, for the first time, generalizes and scales across thousands of shapes and more than ten different data sources. Our essential contribution is NICP, an ICP-style self-supervised task tailored to neural fields. NICP takes a few seconds, is self-supervised, and works out of the box on pre-trained neural fields. We combine it with a localized Neural Field trained on a large MoCap dataset. NSR achieves the state of the art over public benchmarks, and the release of its code and checkpoints will provide the community with …

3D reverse engineering, in which a CAD model is inferred given a 3D scan of a physical object, is a research direction that offers many promising practical applications. This paper proposes TransCAD, an end-to-end transformer-based architecture that predicts the CAD sequence from a point cloud. TransCAD leverages the structure of CAD sequences by using a hierarchical learning strategy. A loop refiner is also introduced to regress sketch primitive parameters. Rigorous experimentation on the DeepCAD and Fusion360 datasets show that TransCAD achieves state-of-the-art results. The result analysis is supported with a proposed metric for CAD sequence, the mean Average Precision of CAD Sequence, that addresses the limitations of existing metrics.

Scanned point clouds are often sparse and incomplete due to the limited field of view of sensing devices, significantly impeding the performance of downstream applications. Therefore, the task of point cloud completion is introduced to obtain a dense and complete point cloud from the incomplete input. The fundamental challenges in tackling this task involve accurately inferring the missing shapes and upsampling them to higher densities. In this paper, we propose a novel approach to address this task, which formulates the completion task as a dual problem: a feature-wise extrapolation problem, where the shape features of the partial point cloud are extrapolated to outlier regions for the recovery of missing portions, and a feature-wise interpolation problem to achieve point cloud upsampling. Based on these, we propose the EINet, a new point cloud completion paradigm with a novel Extrapolation module that can predict the missing shapes for the partial point cloud and a newly designed Interpolation module to upsample the point cloud. Extensive evaluation results demonstrate that EINet achieves compelling performance compared to previous state-of-the-art methods.

Point cloud streaming is increasingly getting popular, evolving into the norm for interactive service delivery and the future Metaverse. However, the substantial volume of data associated with point clouds presents numerous challenges, particularly in terms of high bandwidth consumption and large storage capacity. Despite various solutions proposed thus far, with a focus on point cloud compression, upsampling, and completion, these reconstruction-related methods continue to fall short in delivering high fidelity point cloud output. As a solution, in DiffPMAE, we propose an effective point cloud reconstruction architecture. Inspired by self-supervised learning concepts, we combine Masked Auto-Encoding and Diffusion Model mechanism to remotely reconstruct point cloud data. By the nature of this reconstruction process, DiffPMAE can be extended to many related downstream tasks including point cloud compression, upsampling and completion. Leveraging ShapeNet-55 and ModelNet datasets with over 60000 objects, we validate the performance of DiffPMAE exceeding many state-of-the-art methods in-terms of auto-encoding and downstream tasks considered.

This paper introduces a robust unsupervised SE(3) point cloud registration method that operates without requiring point correspondences. The method frames point clouds as functions in a reproducing kernel Hilbert space (RKHS), leveraging SE(3) equivariant features for direct feature space registration. A novel RKHS distance metric is proposed, offering reliable performance amidst noise, outliers, and asymmetrical data. An unsupervised training approach is introduced to effectively handles limited ground truth data, facilitating adaptation to real datasets. The proposed method outperforms traditional supervised methods in terms of registration accuracy on both synthetic (ModelNet) and real-world (ETH-3D) noisy, outlier-rich datasets, marking the first instance of successful real RGB-D odometry data registration using an equivariant method. The code will be made available upon publication.

Point cloud data, representing the precise 3D layout of the scene, quickly drives the research of 3D object detection. However, the challenge arises due to the rapid iteration of 3D sensors, which leads to significantly different distributions in point clouds. This, in turn, results in subpar performance of 3D cross-sensor object detection. This paper introduces a Cross Mechanism Dataset, named CMD, to support research tackling this challenge. CMD is the first domain adaptation dataset, comprehensively encompassing diverse mechanical sensors and various scenes for 3D object detection. In terms of sensors, CMD includes 32-beam LiDAR, 128-beam LiDAR, solid-state LiDAR, 4D millimeter-wave radar, and cameras, all of which are well-synchronized and calibrated. Regarding the scenes, CMD consists of 50 sequences collocated from different scenarios, ranging from campuses to highways. Furthermore, we validated the effectiveness of various domain adaptation methods in mitigating sensor-based domain differences. We also proposed a DIG method to reduce domain disparities from the perspectives of Density, Intensity, and Geometry, which effectively bridges the domain gap between different sensors. The experimental results on the CMD dataset show that our proposed DIG method outperforms the state-of-the-art techniques, demonstrating the effectiveness of our baseline method. The dataset and the corresponding code are …

Throughout the history of computer vision, while research has explored the integration of images (visual) and point clouds (geometric), many advancements in image and 3D object recognition have tended to process these modalities separately. We aim to bridge this divide by integrating images and point clouds on a unified transformer model. This approach integrates the modality-specific properties of images and point clouds and achieves fundamental downstream tasks in image and 3D object recognition on a unified transformer model by learning visual-geometric representations. In this work, we introduce Formula-Supervised Visual-Geometric Pre-training (FSVGP), a novel synthetic pre-training method that automatically generates aligned synthetic images and point clouds from mathematical formulas. Through cross-modality supervision, we enable supervised pre-training between visual and geometric modalities. FSVGP also reduces reliance on real data collection, cross-modality alignment, and human annotation. Our experimental results show that FSVGP pre-trains more effectively than VisualAtom and PC-FractalDB across six tasks: image and 3D object classification, detection, and segmentation. These achievements demonstrate FSVGP's superior generalization in image and 3D object recognition and underscore the potential of synthetic pre-training in visual-geometric representation learning.

In recent years, robust pre-trained foundation models have been successfully used in many downstream tasks. Here, we would like to use such powerful models to address the problem of few-shot class incremental learning (FSCIL) tasks on 3D point cloud objects. Our approach is to reprogram the well-known CLIP-based foundation model (trained on 2D images and text pairs) for this purpose. The CLIP model works by ingesting 2D images, so to leverage it in our context, we project the 3D object point cloud onto 2D image space to create proper depth maps. For this, prior works consider a fixed and non-trainable set of camera poses. In contrast, we propose to train the network to find a projection that best describes the object and is appropriate for extracting 2D image features from the CLIP vision encoder. Directly using the generated depth map is not suitable for the CLIP model, so we apply the model reprogramming paradigm to the depth map to augment the foreground and background to adapt it. This removes the need for modification or fine-tuning of the foundation model. In the setting we have investigated, we have limited access to data from novel classes, resulting in a problem with overfitting. …

Current feature matching methods prioritize improving modeling capabilities to better align outputs with ground-truth matches, which are the theoretical upper bound on matching results, metaphorically depicted as the “ceiling”. However, these enhancements fail to address the underlying issues that directly hinder ground-truth matches, including the scarcity of matchable points in small scale images, matching conflicts in dense methods, and the keypoint-repeatability reliance in sparse methods. We propose a novel feature matching method named RCM, which Raises the Ceiling of Matching from three aspects. 1) RCM introduces a dynamic view switching mechanism to address the scarcity of matchable points in source images by strategically switching image pairs. 2) RCM proposes a conflict-free coarse matching module, addressing matching conflicts in the target image through a many-to-one matching strategy. 3) By integrating the semi-sparse paradigm and the coarseto- fine architecture, RCM preserves the benefits of both high efficiency and global search, mitigating the reliance on keypoint repeatability. As a result, RCM enables more matchable points in the source image to be matched in an exhaustive and conflict-free manner in the target image, leading to a substantial 260% increase in ground-truth matches. Comprehensive experiments show that RCM exhibits remarkable performance and efficiency in comparison …

We tackle the task of learning dynamic 3D semantic radiance fields given a single monocular video as input. Our learned semantic radiance field captures per-point semantics as well as color and geometric properties for a dynamic 3D scene, enabling the generation of novel views and their corresponding semantics. This enables the segmentation and tracking of a diverse set of 3D semantic entities, specified using a simple and intuitive interface that includes a user click or a text prompt. To this end, we present DGD, a unified 3D representation for both the appearance and semantics of a dynamic 3D scene, building upon the recently proposed dynamic 3D Gaussians representation. Our representation is optimized over time with both color and semantic information. Key to our method is the joint optimization of the appearance and semantic attributes, which jointly affect the geometric properties of the scene. We evaluate our approach in its ability to enable dense semantic 3D object tracking and demonstrate high-quality results that are fast to render, for a diverse set of scenes. Our code will be made available upon acceptance.

Canonical surface mapping is a generalization of keypoint detection with the goal of assigning each pixel of an object to a corresponding point in a 3D template. Popularised by DensePose for the analysis of humans, authors have since attempted to apply the concept to more categories, but with limited success due to the high cost of manual supervision. In this work, we introduce SICK, a method to learn canonical maps without manual supervision which achieves better results than supervised methods for most categories. Our idea is to leverage foundation computer vision models such as DINO and Stable Diffusion that are open-ended and thus possess excellent priors over natural categories. SICK reduces the problem of estimating image-to-template correspondences to predicting image-to-image correspondences using features from the foundation models. The reduction works by matching images of the object to non-photorealistic renders of the template, which emulates the process of collecting manual annotations for this task. These correspondences are then used to supervise high-quality canonical maps for any object of interest. We also show that image generators can further improve the realism of the template views, which provide an additional source of supervision for the model.

We present LineFit, an algorithm that fits line segments from a predicted image gradient map. While existing detectors aim at capturing line segments on line-like structures, our algorithm also seeks to approximate curved shapes. This particularity is interesting for addressing vectorization problems with edge-based representations, after connecting the detected line segments. Our algorithm measures and optimizes the quality of a line segment configuration globally as a point-to-line fitting problem. The quality of configurations is measured through the local fitting error, the completeness over the image gradient map and the capacity to preserve geometric regularities. A key ingredient of our work is an efficient and scalable exploration mechanism that refines an initial configuration by ordered sequences of geometric operations. We show the potential of our algorithm when combined with recent deep image gradient predictions and its competitiveness against existing detectors on different datasets, especially when scenes contain curved objects. We also demonstrate the benefit of our algorithm for polygonalizing objects.

Reconstructing 3D structure and camera motion from images has been a long-standing focus of computer vision research and is commonly referred to as Structure-from-Motion (SfM). Solutions to this problem are categorized into two main approaches: incremental and global. While the most popular systems follow the incremental paradigm due to superior accuracy and robustness, global approaches are drastically more scalable and efficient. We revisit the problem of global SfM and propose GLOMAP as a new general-purpose system. In terms of accuracy and robustness, we achieve results comparable to COLMAP, the most widely used incremental SfM, while being orders of magnitude faster. GLOMAP significantly outperforms state-of-the-art global SfM (Theia, OpenMVG). The code will be made available as open-source.

In this work, we propose the use of Neural Radiance Fields (NeRF) as a scene representation for visual localization. Recently, NeRF has been employed to enhance pose regression and scene coordinate regression models by augmenting the training database, providing auxiliary supervision through rendered images, or serving as an iterative refinement module. We extend its recognized advantages -- its ability to provide a compact scene representation with realistic appearances and accurate geometry -- by exploring the potential of NeRF's internal features in establishing precise 2D-3D matches for localization. To this end, we conduct a comprehensive examination of NeRF's implicit knowledge, acquired through view synthesis, for matching under various conditions. This includes exploring different matching network architectures, extracting encoder features at multiple layers, and varying training configurations. Significantly, we introduce NeRFMatch, an advanced 2D-3D matching function that capitalizes on the internal knowledge of NeRF learned via view synthesis. Our evaluation of NeRFMatch on standard localization benchmarks, within a structure-based pipeline, sets a new state-of-the-art for localization performance on Cambridge Landmarks.

Accurate detection of cephalometric landmarks is crucial for orthodontic diagnosis and treatment planning. Current methods rely on a cascading form of multiple models to achieve higher accuracy, which greatly complicates both training and deployment processes. In this paper, we introduce a novel regression paradigm capable of simultaneously detecting all cephalometric landmarks in high-resolution X-ray images. Our approach only utilizes the encoder module from the transformer to design a dual-encoder architecture, enabling precise detection of cephalometric landmark positions from coarse to fine. Specifically, the entire model architecture comprises three main components: a feature extractor module, a reference encoder module, and a finetune encoder module. These components are respectively responsible for feature extraction and fusion for X-ray images, coarse localization of cephalometric landmark, and fine-tuning of cephalometric landmark positioning. Notably, our framework is fully end-to-end differentiable and innately learns to exploit the interdependencies among cephalometric landmarks. Experiments demonstrate that our method significantly surpasses the current state-of-the-art methods in Mean Radical Error (MRE) and the 2mm Success Detection Rate (SDR) metrics, while also reducing computational resource consumption. Our code will be available soon.

We propose FoundPose, a model-based method for 6D pose estimation of unseen objects from a single RGB image. The method can quickly onboard new objects using their 3D models, without requiring any object- nor task-specific training. In contrast, existing methods typically pre-train on large-scale, task-specific datasets in order to generalize to new objects and to bridge the image-to-model domain gap. We demonstrate that such generalization capabilities can be observed in a recent vision foundation model trained in a self-supervised manner. Specifically, our method estimates the object pose from image-to-model 2D-3D correspondences, which are established by matching patch descriptors from the recent DINOv2 model between the image and pre-rendered object templates. We find that reliable correspondences can be established by kNN matching of patch descriptors from an intermediate DINOv2 layer. Such descriptors carry stronger positional information than descriptors from the last layer, and we show their importance when semantic information is ambiguous due to object symmetries or a lack of texture. To avoid establishing correspondences against all object templates, we develop an efficient template retrieval approach that integrates the patch descriptors into the bag-of-words representation, and can promptly propose a handful of similarly looking templates. Additionally, we apply featuremetric alignment to …

Salient Object Ranking (SOR) aims to study how humans shift their attention among various objects within a scene. Previous works attempt to excavate explicit visual saliency cues, e.g., spatial frequency and semantic context, to tackle this challenge. However, these visual saliency cues may fall short in handling real-world scenarios, which often involve various human activities and interactions. We observe that human observers' attention can be reflexively guided by the poses and gestures of the people in the scene, which indicate their activities. For example, observers tend to shift their attention to follow others' head orientation or running/walking direction to anticipate what will happen. Inspired by this observation, we propose to explore the human skeletal pose to deeply understand high-level interactions between human participants and their surroundings for robust salient object ranking. Specifically, we propose PoseSOR, a human pose-aware SOR model for the SOR task, with two novel modules: 1) a Pose-Aware Interaction (PAI) Module to integrate human pose knowledge into salient object queries for learning high-level interactions, and 2) a Pose-Driven Ranking (PDR) Module to apply pose knowledge as directional cues to help predict where human attention will shift to. To our knowledge, our approach is the first to explore …

Traditional 2D pose estimation models are limited by their category-specific design, making them suitable only for predefined object categories. This restriction becomes particularly challenging when dealing with novel objects due to the lack of relevant training data. To address this limitation, category-agnostic pose estimation (CAPE) was introduced. CAPE aims to enable keypoint localization for arbitrary object categories using a few-shot single model, requiring minimal support images with annotated keypoints. We present a significant departure from conventional CAPE techniques, which treat keypoints as isolated entities, by treating the input pose data as a graph. We leverage the inherent geometrical relations between keypoints through a graph-based network to break symmetry, preserve structure, and better handle occlusions. We validate our approach on the MP-100 benchmark, a comprehensive dataset comprising over 20,000 images spanning over 100 categories. Our solution boosts performance by 0.98% under a 1-shot setting, achieving a new state-of-the-art for CAPE. Additionally, we enhance the dataset with skeleton annotations. Our code and data are publicly available.

In this study, we introduce the 3D space attention module (3DSA) as a novel approach to address the drawback of multi-view 3D human pose estimation methods, which fail to recognize the object's significance from diverse viewpoints. Specifically, we utilize the 3D space subdivision algorithm to divide the feature volume into multiple regions. Predicted 3D space attention scores are assigned to the different regions to construct the feature volume with space attention.The purpose of the 3D space attention module is to distinguish the significance of individual regions within the feature volume by applying weighted attention adjustments derived from corresponding viewpoints. We conduct experiments on existing voxel-based methods, VoxelPose and Faster VoxelPose.By incorporating the space attention module, both achieve state-of-the-art performance on the Panoptic 3D Human Pose Estimation.

WiFi-based human pose estimation (HPE) has emerged as a promising alternative to conventional vision-based techniques, yet faces the high computational cost hindering its widespread adoption. This paper introduces a novel HPE-Li approach that harnesses multi-modal sensors (e.g. camera and WiFi) to generate accurate 3D skeletal in HPE. We then develop an efficient deep neural network to process raw WiFi signals. Our model incorporates a distinctive multi-branch convolutional neural network (CNN) empowered by a selective kernel attention (SKA) mechanism. Unlike standard CNNs with fixed receptive fields, the SKA mechanism is capable of dynamically adjusting kernel sizes according to input data characteristics, enhancing adaptability without increasing complexity. Extensive experiments conducted on two MM-Fi and WiPose datasets underscore the superiority of our method over state-of-the-art approaches, while ensuring minimal computational overhead, rendering it highly suitable for large-scale scenarios.

The extraction of keypoint positions from input hand frames, known as 3D hand pose estimation, is crucial for various human-computer interaction applications. However, current approaches often struggle with the dynamic nature of self-occlusion of hands and intra-occlusion with interacting objects. To address this challenge, this paper proposes the Denoising Adaptive Graph Transformer, HandDAGT, for hand pose estimation. The proposed HandDAGT leverages a transformer structure to thoroughly explore effective geometric features from input patches. Additionally, it incorporates a novel attention mechanism to adaptively weigh the contribution of kinematic correspondence and local geometric features for the estimation of specific keypoints. This attribute enables the model to adaptively employ kinematic and local information based on the occlusion situation, enhancing its robustness and accuracy. Furthermore, we introduce a novel denoising training strategy aimed at improving the model's robust performance in the face of occlusion challenges. Experimental results show that the proposed model significantly outperforms the existing methods on four challenging hand pose benchmark datasets. Codes and pre-trained models are publicly available at https://anonymous.4open.science/r/HandDAGT-C313.

Estimating human and camera trajectories with accurate scale in the world coordinate system from a monocular video is a highly desirable yet challenging and ill-posed problem. In this study, we aim to recover expressive parametric human models (i.e., SMPL-X) and corresponding camera poses jointly, by leveraging the synergy between three critical players: the world, the human, and the camera. Our approach is founded on two key observations. Firstly, camera-frame SMPL-X estimation methods readily recover absolute human depth. Secondly, human motions inherently provide absolute spatial cues. By integrating these insights, we introduce a novel framework, referred to as WHAC, to facilitate world-grounded expressive human pose and shape estimation (EHPS) alongside camera pose estimation, without relying on traditional optimization techniques. Additionally, we present a new synthetic dataset, WHAC-A-Mole, which includes accurately annotated humans and cameras, and features diverse interactive human motions as well as realistic camera trajectories. Extensive experiments on both standard and newly established benchmarks highlight the superiority and efficacy of our framework. We will make the code and dataset publicly available at https://wqyin.github.io/projects/WHAC/.

Accurate tracking of a user’s body pose while wearing a virtual reality (VR), augmented reality (AR) or mixed reality (MR) headset is a prerequisite for authentic self-expression, natural social presence, and intuitive user interfaces. Existing body tracking approaches on VR/AR devices are either under-constrained, e.g., attempting to infer full body pose from only headset and controller pose, or require impractical hardware setups that place cameras far from a user’s face to improve body visibility. In this paper, we present the first controllerless egocentric body tracking solution that runs on an actual VR device using the same cameras that are used for SLAM tracking. We propose a novel egocentric tracking architecture that models the temporal history of body motion using multi-view latent features. Furthermore, we release the first large-scale real-image dataset for egocentric body tracking, EgoBody3M, with a realistic VR headset configuration and diverse subjects and motions. Benchmarks on the dataset shows that our approach outperforms other state-of-the-art methods in both accuracy and smoothness of the resulting motion. We perform ablation studies on our model choices and demonstrate the method running in realtime on a VR headset. Our dataset with more than 30 hours of recordings and 3 million frames will …

Temporal dependencies are essential in 3D human pose estimation to mitigate depth ambiguity. Previous methods typically use a fixed-length sliding window to capture these dependencies. However, they treat past and future frames equally, ignoring the fact that relying on too many future frames increases the inference latency. In this paper, we present a 3D human pose estimation model based on Retentive Networks (RetNet) that incorporates temporal information by utilizing a large number of past frames and a few future frames. The Non-Causal RetNet (NC-RetNet) is designed to allow the originally causal RetNet to be aware of future information. Additionally, we propose a knowledge transfer strategy, i.e., training the model with a larger chunk size and using a smaller chunk size during inference, to reduce latency while maintaining comparable accuracy. Extensive experiments have been conducted on the Human3.6M and MPI-INF-3DHP datasets, and the results demonstrate that our method achieves state-of-the-art performance.

Enabling robotic manipulation that generalizes to out-of-distribution scenes is a crucial step toward open-world embodied intelligence. For human beings, this ability is rooted in the understanding of semantic correspondence among objects, which helps to naturally transfer the interaction experience of familiar objects to novel ones. Although robots lack such a reservoir of interaction experience, the vast availability of human videos on the Internet may serve as a resource, from which we extract an affordance memory of contact points. Inspired by the natural way humans think, We propose Robo-ABC: when confronted with unfamiliar objects that require generalization, the robot can acquire affordance by retrieving objects that share visual and semantic similarities from the memory, then mapping the contact points of the retrieved objects to the new object. While such correspondence may present formidable challenges at first glance, recent research finds it naturally arises from pre-trained diffusion models, enabling affordance mapping even across disparate categories. Through the Robo-ABC framework, robots can generalize to manipulate out-of-category objects in a zero-shot manner without any manual annotation, additional training, part segmentation, pre-coded knowledge, or viewpoint restrictions. Quantitatively, Robo-ABC significantly enhances the accuracy of visual affordance inference by a large margin of 28.7% compared to state-of-the-art …

Modern smartphone camera quality heavily relies on the image signal processor (ISP) to enhance captured raw images, utilizing carefully designed modules to produce final output images encoded in a standard color space (e.g., sRGB). Neural-based end-to-end learnable ISPs offer promising advancements, potentially replacing traditional ISPs with their ability to adapt without requiring extensive tuning for each new camera model, as is often the case for nearly every module in traditional ISPs. However, the key challenge with the recent learning-based ISPs is the urge to collect large paired datasets for each distinct camera model due to the influence of intrinsic camera characteristics on the formation of input raw images. This paper tackles this challenge by introducing a novel method for unpaired learning of raw-to-raw translation across diverse cameras. Specifically, we propose Rawformer, an unsupervised Transformer-based encoder-decoder method for raw-to-raw translation. It accurately maps raw images captured by a certain camera to the target camera, facilitating the generalization of learnable ISPs to new unseen cameras. Our method demonstrates superior performance on real camera datasets, achieving higher accuracy compared to previous state-of-the-art techniques, and preserving a more robust correlation between the original and translated raw images. Code and models will be publicly available …

We introduce the Reality-linked 3D Scenes (R3DS) dataset of synthetic 3D scenes mirroring the real-world scene arrangements from Matterport3D panoramas. Compared to prior work, R3DS has more complete and densely populated scenes with objects linked to real-world observations in panoramas. R3DS also provides an object support hierarchy, and matching object sets (e.g., same chairs around a dining table) for each scene. Overall, R3DS contains 19K objects represented by 3,784 distinct CAD models from over 100 object categories. We demonstrate the effectiveness of R3DS on the Panoramic Scene Understanding task. We find that: 1) training on R3DS enables better generalization; 2) support relation prediction trained with R3DS improves performance compared to heuristically calculated support; and 3) R3DS offers a challenging benchmark for future work on panoramic scene understanding.

Self-supervised monocular depth estimation has gathered notable interest since it can liberate training from dependency on depth annotations. In monocular video training case, recent methods only conduct view synthesis between existing camera views, leading to insufficient guidance. To tackle this, we try to synthesize more virtual camera views by flow-based video frame interpolation (VFI), termed as temporal augmentation. For multi-frame inference, to sidestep the problem of dynamic objects encountered by explicit geometry-based methods like ManyDepth, we return to the feature fusion paradigm and design a VFI-assisted multi-frame fusion module to align and aggregate multi-frame features, using motion and occlusion information obtained by the flow-based VFI model. Finally, we construct a unified self-supervised learning framework, named Mono-ViFI, to bilaterally connect single- and multi-frame depth. In this framework, spatial data augmentation through image affine transformation is incorporated for data diversity, along with a triplet depth consistency loss for regularization. The single- and multi-frame models can share weights, making our framework compact and memory-efficient. Extensive experiments on several datasets demonstrate that our method can bring significant improvements to current advanced architectures without increasing inference complexity.

In this paper, we propose a novel video depth estimation approach, FutureDepth, which enables the model to implicitly leverage multi-frame and motion cues to improve depth estimation by making it learn to predict the future at training. More specifically, we propose a future prediction network, F-Net, which takes the features of multiple consecutive frames and is trained to predict multi-frame features one time step ahead iteratively. In this way, F-Net learns the underlying motion and correspondence information, and we incorporate its features into the depth decoding process. Additionally, to enrich the learning of multi-frame correspondence cues, we further leverage a reconstruction network, R-Net, which is trained via adaptively masked auto-encoding of multi-frame feature volumes. At inference time, both F-Net and R-Net are used to produce queries to work with the depth decoder, as well as a final refinement network. Through extensive experiments on several benchmarks, i.e., NYUDv2, KITTI, DDAD, and Sintel, which cover indoor, driving, and open-domain scenarios, we show that FutureDepth significantly improves upon baseline models, outperforms existing video depth estimation methods, and sets new state-of-the-art (SOTA) accuracy. Furthermore, FutureDepth is more efficient than existing SOTA video depth estimation models and has similar latencies when comparing to monocular models.

Perspective distortion (PD) causes unprecedented changes in shape, size, orientation, angles, and other spatial relationships of visual concepts in images. Precisely estimating camera intrinsic and extrinsic parameters is a challenging task that prevents synthesizing perspective distortion. Non-availability of dedicated training data poses a critical barrier to developing robust computer vision methods. Additionally, distortion correction methods make other computer vision tasks a multi-step approach and lack performance. In this work, we propose mitigating perspective distortion (MPD) by employing a fine-grained parameter control on a specific family of Möbius transform to model real-world distortion without estimating camera intrinsic and extrinsic parameters and without the need for actual distorted data. Also, we present a dedicated perspectively distorted benchmark dataset, ImageNet-PD, to benchmark the robustness of deep learning models against this new dataset. The proposed method outperforms on existing benchmarks, ImageNet-E and ImageNet-X. Additionally, it significantly improves performance on ImageNet-PD while consistently performing on standard data distribution. Further, our method shows improved performance on three PD-affected real-world applications: crowd counting, fisheye image recognition, and person re-identification. We will release source code, dataset, and models for foster further research.

Data-driven visual-inertial odometry (VIO) has received highlights for its performance since VIOs are a crucial compartment in autonomous robots. However, their deployment on resource-constrained devices is non-trivial since large network parameters should be accommodated in the device memory. Furthermore, these networks may risk failure post-deployment due to environmental distribution shifts at test time. In light of this, we propose UL-VIO -- an ultra-lightweight (<1M) VIO network capable of test-time adaptation (TTA) based on visual-inertial consistency. Specifically, we perform model compression to the network while preserving the low-level encoder part, including all BatchNorm parameters for resource-efficient test-time adaptation. It achieves 36X smaller network size than state-of-the-art with a minute increase in error -- 1% on the KITTI dataset. For test-time adaptation, we propose to use the inertia-referred network outputs as pseudo labels and update the BatchNorm parameter for lightweight yet effective adaptation. To the best of our knowledge, this is the first work to perform noise-robust TTA on VIO. Experimental results on the KITTI, EuRoC, and Marulan datasets demonstrate the effectiveness of our resource-efficient adaptation method under diverse TTA scenarios with dynamic domain shifts.

Camera-based Bird's-Eye-View (BEV) perception often struggles between adopting 3D-to-2D or 2D-to-3D view transformation (VT). The 3D-to-2D VT typically employs resource-intensive Transformer to establish robust correspondences between 3D and 2D feature, while the 2D-to-3D VT utilizes the Lift-Splat-Shoot (LSS) pipeline for real-time application, potentially missing distant information. To address these limitations, we propose DualBEV, a unified framework that utilizes a shared feature transformation incorporating three probabilistic measurements for both strategies. By considering dual-view correspondences in one-stage, DualBEV effectively bridges the gap between these strategies, harnessing their individual strengths. Our method achieves state-of-the-art performance without Transformer, delivering comparable efficiency to the LSS approach, with 55.2% mAP and 63.4% NDS on the nuScenes test set. Code will be released.

Three-dimensional perception from multi-view cameras is a crucial component in autonomous driving systems, which involves multiple tasks like 3D object detection and bird's-eye-view (BEV) semantic segmentation. To improve perception precision, large image encoders, high-resolution images, and long-term temporal inputs have been adopted in recent 3D perception models, bringing remarkable performance gains. However, these techniques are often incompatible in training and inference scenarios due to computational resource constraints. Besides, modern autonomous driving systems prefer to adopt an end-to-end framework for multi-task 3D perception, which can simplify the overall system architecture and reduce the implementation complexity. However, conflict between tasks often arises when optimizing multiple tasks jointly within an end-to-end 3D perception model. To alleviate these issues, we present an end-to-end framework named HENet for multi-task 3D perception in this paper. Specifically, we propose a hybrid image encoding network, using a large image encoder for short-term frames and a small image encoder for long-term temporal frames. Then, we introduce a temporal feature integration module based on the attention mechanism to fuse the features of different frames extracted by the two aforementioned hybrid image encoders. Finally, according to the characteristics of each perception task, we utilize BEV features of different grid sizes, independent …

The field of autonomous driving has attracted considerable interest in approaches that directly infer 3D objects in the Bird's Eye View (BEV) from multiple cameras. Some attempts have also explored utilizing 2D detectors from single images to enhance the performance of 3D detection. However, these approaches rely on a two-stage process with separate detectors, where the 2D detection results are utilized only once for token selection or query initialization. In this paper, we present a single model termed SimPB, which Simultaneously detects 2D objects in the Perspective view and 3D objects in the BEV space from multiple cameras. To achieve this, we introduce a hybrid decoder consisting of several multi-view 2D decoder layers and several 3D decoder layers, specifically designed for their respective detection tasks. A Dynamic Query Allocation module and an Adaptive Query Aggregation module are proposed to continuously update and refine the interaction between 2D and 3D results, in a cyclic 3D-2D-3D manner. Additionally, Query-group Attention is utilized to strengthen the interaction among 2D queries within each camera group. In the experiments, we evaluate our method on the nuScenes dataset and demonstrate promising results for both 2D and 3D detection tasks. We will make the code publicly available.

Weakly supervised 3D object detection aims to learn a 3D detector with lower annotation cost, e.g., 2D labels. Unlike prior work which still relies on few accurate 3D annotations, we propose a framework to study how to leverage constraints between 2D and 3D domains without requiring any 3D labels. Specifically, we employ visual data from three perspectives to establish connections between 2D and 3D domains. First, we design a feature-level constraint to align LiDAR and image features based on object-aware regions. Second, the output-level constraint is developed to enforce the overlap between 2D and projected 3D box estimations. Finally, the training-level constraint is utilized by producing accurate and consistent 3D pseudo-labels that align with the visual data. We conduct extensive experiments on the KITTI dataset to validate the effectiveness of the proposed three constraints. Without using any 3D labels, our method achieves favorable performance against state-of-the-art approaches and is competitive with the method that uses 500-frame 3D annotations. Code and models will be made publicly available.

Popular representation learning methods encourage feature invariance under transformations applied at the input. However, in 3D perception tasks like object localization and segmentation, outputs are naturally equivariant to some transformations, such as rotation. Using pre-training loss functions that encourage equivariance of features under certain transformations provides a strong self-supervision signal while also retaining information of geometric relationships between transformed feature representations. This can enable improved performance in downstream tasks that are equivariant to such transformations. In this paper, we propose a spatio-temporal equivariant learning framework by considering both spatial and temporal augmentations jointly. Our experiments show that the best performance arises with a pre-training approach that encourages equivariance to translation, scaling, and flip, rotation and scene flow. For spatial augmentations, we find that depending on the transformation, either a contrastive objective or an equivariance-by-classification objective yields best results. To leverage real-world object deformations and motion, we consider sequential LiDAR scene pairs and develop a novel 3D scene flow-based equivariance objective that leads to improved performance overall. We show our pre-training method for 3D object detection which outperforms existing equivariant and invariant approaches in many settings.

LiDAR-based 3D object detection models show remarkable performance, however their effectiveness diminishes in adverse weather. On the other hand, 4D radar exhibits strengths in adverse weather but faces limitations in standalone use. While fusing LiDAR and 4D radar seems to be the most intuitive approach, this method comes with limitations, including increased computational load due to radar pre-processing, situational constraints when both domain information is present, and the potential loss of sensor advantages through joint optimization. In this paper, we propose a novel LiDAR-only-based 3D object detection framework that works robustly in all-weather (normal and adverse) conditions. Specifically, we first propose 4D radar-based 3D prompt learning to inject auxiliary radar information into a LiDAR-based pre-trained 3D detection model while preserving the precise geometry capabilities of LiDAR. Subsequently, using the preceding model as a teacher, we distill weather-insensitive features and responses into a LiDAR-only student model through our four levels of inter-/intra-modal knowledge distillation. Extensive experiments demonstrate that our prompt learning effectively integrates the strengths of LiDAR and 4D radar, and our LiDAR-only student model even surpasses the detection performance of teacher and state-of-the-art models under various weather. Code will be released.

Multimodal sensor fusion is an essential capability for autonomous robots, enabling object detection and decision-making in the presence of failing or uncertain inputs. While recent fusion methods excel in normal environmental conditions, these methods fail in adverse weather conditions, e.g., heavy fog, snow, or obstructions due to soiling. To address these challenges, we introduce a novel multi sensor fusion approach tailored for adverse weather conditions. In addition to RGB and LiDAR sensors employed in recent autonomous driving literature, our sensor fusion stack is capable of learning from NIR Gated camera and radar modalities to tackle low light and adverse weather conditions. We propose to fuse multimodal sensor data through attentive, depth-based blending schemes, with learned refinement in the Bird's Eye View (BEV) domain to combine image and range features. Our detections are predicted by a transformer decoder that weights modalities based on distance and visibility. We validate that our method improves the reliability of multimodal sensor fusion in autonomous vehicles under challenging weather conditions, bridging the gap between ideal conditions and real-world edge cases and improving average precision by 17.6 AP points to the second best method in the pedestrian class in long range dense fog conditions.

Collaborative perception has received widespread attention recently since it enhances the perception ability of autonomous vehicles via inter-agent information sharing. However, the performance of existing systems is hindered by the unavoidable collaboration noises, which induce feature-level spatial misalignment over the collaborator-shared information. In this paper, we propose a model-agnostic and lightweight plugin to mitigate the feature-level misalignment issue, called dynamic feature alignment (NEAT). The merits of the NEAT plugin are threefold. First, we introduce an importance-guided query proposal to predict potential foreground regions with space-channel semantics and exclude environmental redundancies. On this basis, a deformable feature alignment is presented to explicitly align the collaborator-shared features through query-aware spatial associations, aggregating multi-grained visual clues with corrective mismatch properties. Ultimately, we perform a region cross-attention reinforcement to facilitate aligned representation diffusion and achieve global feature semantic enhancement. NEAT can be readily inserted into existing collaborative perception procedures and significantly improves the robustness of vanilla baselines against pose errors and transmission delay. Extensive experiments on four collaborative 3D object detection datasets under noisy settings confirm that NEAT provides consistent gains for most methods with distinct structures.

Real-world aerial scene understanding is limited by a lack of datasets that contain densely annotated images curated under a diverse set of conditions. Due to inherent challenges in obtaining such images in controlled real-world settings, we present SkyScenes, a synthetic dataset of densely annotated aerial images captured from Unmanned Aerial Vehicle (UAV) perspectives. We carefully curate SkyScenes images from CARLA to comprehensively capture diversity across layouts (urban and rural maps), weather conditions, times of day, pitch angles and altitudes with corresponding semantic, instance and depth annotations. Through our experiments using SkyScenes, we show that (1) models trained on SkyScenes generalize well to different real-world scenarios, (2) augmenting training on real images with SkyScenes data can improve real-world performance, (3) controlled variations in SkyScenes can offer insights into how models respond to changes in viewpoint conditions, and (4) incorporating additional sensor modalities (depth) can improve aerial scene understanding.

With the surge in autonomous driving technologies, the reliance on comprehensive and high-definition bird's-eye-view (BEV) representations has become paramount. This burgeoning need underscores the demand for extensive multi-view video datasets, meticulously annotated to facilitate advanced research and development. Nonetheless, the acquisition of such datasets is impeded by prohibitive costs associated with data collection and annotation. There are two challenges when synthesizing multi-view videos given a 3D layout: Generating multi-view videos involves handling both view and temporal dimensions. How to generate videos while ensuring cross-view consistency and cross-frame consistency? 2) How to ensure the precision of layout control and the quality of the generated instances? Addressing this critical bottleneck, we introduce a novel spatial-temporal consistent diffusion framework, DrivingDiffusion, engineered to synthesize realistic multi-view videos governed by 3D spatial layouts. DrivingDiffusion adeptly navigates the dual challenges of maintaining cross-view and cross-frame coherence, along with meeting the exacting standards of layout fidelity and visual quality. The framework operates through a tripartite methodology: initiating with the generation of multi-view single-frame images, followed by the synthesis of single-view videos across multiple cameras, and culminating with a post-processing phase. We corroborate the efficacy of DrivingDiffusion through rigorous quantitative and qualitative evaluations, demonstrating its potential to significantly …

Vehicle trajectory prediction has increasingly relied on data-driven solutions, but their ability to scale to different data domains and the impact of larger dataset sizes on their generalization remain under-explored. While these questions can be studied by employing multiple datasets, it is challenging due to several discrepancies, e.g., in data formats, map resolution, and semantic annotation types. To address these challenges, we introduce UniTraj, a comprehensive framework that unifies various datasets, models, and evaluation criteria, presenting new opportunities for the vehicle trajectory prediction field. In particular, using UniTraj, we conduct extensive experiments and find that model performance significantly drops when transferred to other datasets. However, enlarging data size and diversity can substantially improve performance, leading to a new state-of-the-art result for the nuScenes dataset. We provide insights into dataset characteristics to explain these findings. We will release the framework to support further research.

Generating 3D vehicle assets from in-the-wild observations is crucial to autonomous driving. Existing image-to-3D methods cannot well address this problem because they learn generation merely from image RGB information without a deeper understanding of in-the-wild vehicles (such as car models, manufacturers, etc). This leads to their poor zero-shot prediction capability to handle real-world observations with occlusion or tricky viewing angles. To solve this problem, in this work, we propose VQA-Diff, a novel framework that leverages in-the-wild vehicle images to create photorealistic 3D vehicle assets for autonomous driving. VQA-Diff exploits the real-world knowledge inherited from the Large Language Model in the Visual Question Answering (VQA) model for robust zero-shot prediction and the rich image prior knowledge in Diffusion Models for structure and appearance generation. In particular, we utilize a multi-expert Diffusion Models strategy to generate the structure information and employ a subject-driven structure-controlled generation mechanism to model appearance information. As a result, without the necessity to learn from a large-scale image-to-3D vehicle dataset collected from the real world, VQA-Diff still has a robust zero-shot image-to-novel-view generation ability. We conduct experiments on various datasets, including Pascal 3D+, Waymo, and Objaverse, to demonstrate that VQA-Diff outperforms existing state-of-the-art methods both qualitatively and quantitatively.

Existing solutions for 3D semantic occupancy prediction typically treat the task as a one-shot 3D voxel-wise segmentation perception problem. These discriminative methods focus on learning the mapping between the inputs and occupancy map in a single step, lacking the ability to gradually refine the occupancy map and the reasonable scene imaginative capacity to complete the local regions somewhere. In this paper, we introduce OccGen, a simple yet powerful generative perception model for the task of 3D semantic occupancy prediction. OccGen adopts a noise-to-occupancy'' generative paradigm, progressively inferring and refining the occupancy map by predicting and eliminating noise originating from a random Gaussian distribution. OccGen consists of two main components: a conditional encoder that is capable of processing multi-modal inputs, and a progressive refinement decoder that applies diffusion denoising using the multi-modal features as conditions. A key insight of this generative pipeline is that the diffusion denoising process is naturally able to model the coarse-to-fine refinement of the dense 3D occupancy map, therefore producing more detailed predictions. Extensive experiments on several occupancy benchmarks demonstrate the effectiveness of the proposed method compared to the state-of-the-art methods. For instance, OccGen relatively enhances the mIoU by 9.5%, 6.3%, and 13.3% on nuScenes-Occupancy dataset under …

This paper introduces the Stream Query Denoising (SQD) strategy, a novel and general approach for high-definition map (HD-map) construction. SQD is designed to improve the modeling capability of map elements by learning temporal consistency. Specifically, SQD involves the process of denoising the queries, which are generated by the noised ground truth of the previous frame. This process aims to reconstruct the ground truth of the current frame during training. Our method can be applied to both static and temporal methods, showing the great effectiveness of SQD strategy. Extensive experiments on nuScenes and Argoverse2 show that our framework achieves superior performance, compared to other existing methods across all settings.

Understanding road geometry is a critical component of the autonomous vehicle (AV) stack. While high-definition (HD) maps can readily provide such information, they suffer from high labeling and maintenance costs. Accordingly, many recent works have proposed methods for estimating HD maps online from sensor data. The vast majority of recent approaches encode multi-camera observations into an intermediate representation, e.g., a bird's eye view (BEV) grid, and produce vector map elements via a decoder. While this architecture is performant, it decimates much of the information encoded in the intermediate representation, preventing downstream tasks (e.g., behavior prediction) from leveraging them. In this work, we propose exposing the rich internal features of online map estimation methods and show how they enable more tightly integrating online mapping with trajectory forecasting. In doing so, we find that directly accessing internal BEV features yields up to 73% faster inference speeds and up to 29% more accurate predictions on the real-world nuScenes dataset.

Prior works have addressed the problem of driver intention prediction (DIP) by identifying maneuvers after their onset. On the other hand, early anticipation is equally important in scenarios that demand a preemptive response before a maneuver begins. However, there is no prior work aimed at addressing the problem of driver action anticipation before the onset of the maneuver, limiting the ability of the advanced driver assistance system (ADAS) for early maneuver anticipation. In this work, we introduce Anticipating Driving Maneuvers (ADM), a new task that enables driver action anticipation before the onset of the maneuver. To initiate research in ADM task, we curate Driving Action Anticipation Dataset, DAAD, that is multi-view: in- and out-cabin views in dense and heterogeneous scenarios, and multimodal: egocentric view and gaze information. The dataset captures sequences both before the initiation and during the execution of a maneuver. During dataset collection, we also ensure to capture wide diversity in traffic scenarios, weather and illumination, and driveway conditions. Next, we propose a strong baseline based on a transformer architecture to effectively model multiple views and modalities over longer video lengths. We benchmark the existing DIP methods on DAAD and related datasets. Finally, we perform an ablation study …

Human trajectory prediction is typically posed as a zero-shot generalization problem: a predictor is learnt on a dataset of human motion in training scenes, and then deployed on unseen test scenes. While this paradigm has yielded tremendous progress, it fundamentally assumes that trends in human behavior within the deployment scene are constant over time. As such, current prediction models are unable to adapt to scene-specific transient human behaviors, such as crowds temporarily gathering to see buskers, pedestrians hurrying through the rain and avoiding puddles, or a protest breaking out. We formalize the problem of scene-specific adaptive trajectory prediction and propose a new adaptation approach inspired by prompt tuning called latent corridors. By augmenting the input of any pre-trained human trajectory predictor with learnable image prompts, the predictor can improve in the deployment scene by inferring trends from extremely small amounts of new data (e.g., 2 humans observed for 30 seconds). With less than 0.1% additional model parameters, we see up to 23.9% ADE improvement in MOTSynth simulated data and 16.4% ADE in MOT and Wildtrack real pedestrian data. Qualitatively, we observe that latent corridors imbue predictors with an awareness of scene geometry and scene-specific human behaviors that non-adaptive predictors struggle …

Modeling the trajectories of intelligent vehicles is an essential component of a traffic-simulating system. However, such trajectory predictors are typically trained to imitate the movements of human drivers. The imitation models often fall short of capturing safety-critical events residing in the long-tail end of the data distribution, especially under complex environments involving multiple drivers. In this paper, we propose a game-theoretic perspective to resolve this challenge by modeling the competitive interactions of vehicles in a general-sum Markov game and characterizing these safety-critical events with the correlated equilibrium. To achieve this goal, we pretrain a generative world model to predict the environmental dynamics of self-driving scenarios. Based on this world model, we probe the action predictor for identifying the Coarse Correlated Equilibrium (CCE) by incorporating both optimistic Bellman update and magnetic mirror descent into the objective function of the Multi-Agent Reinforcement Learning (MARL) algorithm. We conduct extensive experiments to demonstrate our algorithm outperforms other baselines in terms of efficiently closing the CCE-gap and generating meaningful trajectories under competitive autonomous driving environments.

Weather forecasting requires both deterministic outcomes for immediate decision-making and probabilistic results for assessing uncertainties. However, deterministic models may not fully capture the spectrum of weather possibilities, and probabilistic forecasting can lack the precision needed for specific planning, presenting significant challenges as the field aims for enhance accuracy and reliability. In this paper, we propose the Deterministic Guidance-based Diffusion Model (DGDM) to exploit the benefits of both deterministic and probabilistic weather forecasting models. DGDM integrates a deterministic branch and a diffusion model as a probabilistic branch to improve forecasting accuracy while providing probabilistic forecasting. In addition, we introduce a sequential variance schedule that predicts from the near future to the distant future. Moreover, we present a truncated diffusion by using the result of the deterministic branch to truncate the reverse process of the diffusion model to control uncertainties. We conduct extensive analyses of DGDM on the Moving MNIST. Furthermore, we evaluate the effectiveness of DGDM on the Pacific Northwest Windstorm (PNW)-Typhoon satellite dataset for regional extreme weather forecasting, as well as on the WeatherBench dataset for global weather forecasting dataset. Experimental results show that DGDM achieves state-of-the-art performance not only in global forecasting but also in regional forecasting scenarios. The …

Current optical flow and point-tracking methods rely heavily on synthetic datasets. Event cameras are novel vision sensors with advantages in challenging visual conditions, but state-of-the-art frame-based methods cannot be easily adapted to event data due to the limitations of current event simulators. We introduce a novel self-supervised loss combining the Contrast Maximization framework with a non-linear motion prior in the form of pixel-level trajectories and propose an efficient solution to solve the high-dimensional assignment problem between non-linear trajectories and events. Their effectiveness is demonstrated in two scenarios: In dense continuous-time motion estimation, our method improves the zero-shot performance of a synthetically trained model on the real-world dataset EVIMO2 by 29%. In optical flow estimation, our method elevates a simple UNet to achieve state-of-the-art performance among self-supervised methods on the DSEC optical flow benchmark.

Event cameras are dynamic vision sensors inspired by the biological retina, characterized by their high dynamic range, high temporal resolution, and low power consumption. These features make them capable of perceiving 3D environments even in extreme conditions. Event data is continuous across the time dimension, which allows a detailed description of each pixel's movements. To fully utilize the temporally dense and continuous nature of event cameras, we propose a novel temporal event stereo, a framework that continuously uses information from previous time steps. This is accomplished through the simultaneous training of an event stereo matching network alongside stereoscopic flow, a new concept that captures all pixel movements from stereo cameras. Since obtaining ground truth for optical flow during training is challenging, we propose a method that uses only disparity maps to train the stereoscopic flow. Ultimately, we enhance the performance of event-based stereo matching by temporally aggregating information using the flows. We have achieved state-of-the-art performance on the MVSEC and the DSEC dataset. Our method is computationally efficient as it stacks previous information in a cascading manner.

Event cameras are neuromorphic image sensors that respond to per-pixel brightness changes, producing a stream of asynchronous and spatially sparse events. Currently, the most successful algorithms for event cameras convert batches of events into dense image-like representations that are synchronously processed by deep learning models of frame-based computer vision. These methods discard the inherent properties of events, leading to high latency and computational costs. Following a recent line of works, we propose a model for efficient asynchronous event processing that exploits sparsity. We design the Fully Asynchronous, Recurrent and Sparse Event-Based CNN (FARSE-CNN), a novel multi-layered architecture which combines the mechanisms of recurrent and convolutional neural networks. To build efficient deep networks, we propose compression modules that allow to learn hierarchical features both in space and time. We theoretically derive the complexity of all components in our architecture, and experimentally validate our method on tasks for object recognition, object detection and gesture recognition. FARSE-CNN achieves similar or better performance than the state-of-the-art among asynchronous methods, with low computational complexity and without relying on a fixed-length history of events. Our code will be released on GitHub.

Introducing event cameras into video super-resolution (VSR) shows great promise. In practice, however, integrating event data as a new modality necessitates a laborious model architecture design. This not only consumes substantial time and effort but also disregards valuable insights from successful existing VSR models. Furthermore, the resource-intensive process of retraining these newly designed structures exacerbates the challenge. In this paper, inspired by recent success of parameter-efficient tuning in reducing the number of trainable parameters of a pre-trained model for downstream tasks, we introduce the Event AdapTER (EATER) for VSR. EATER efficiently utilizes pre-trained VSR model knowledge at the feature level through two lightweight and trainable components: the event-adapted alignment (EAA) unit and the event-adapted fusion (EAF) unit. The EAA unit aligns multiple frames based on the event stream in a coarse-to-fine manner, while the EAF unit efficiently fuses frames with the event stream through a multi-scaled design. Thanks to both units, EATER outperforms the full fine-tuning paradigm. Comprehensive experiments demonstrate the effectiveness of EATER, achieving superior results with parameter efficiency.

In this paper, we tackle the problem of scene-aware 3D human motion forecasting. A key challenge of this task is to predict future human motions that are consistent with the scene by modeling the human-scene interactions. While recent works have demonstrated that explicit constraints on human-scene interactions can prevent the occurrence of ghost motion, they only provide constraints on partial human motion e.g., the global motion of the human or a few joints contacting the scene, leaving the rest of the motion unconstrained. To address this limitation, we propose to model the human-scene interaction with the mutual distance between the human body and the scene. Such mutual distances constrain both the local and global human motion, resulting in a whole-body motion constrained prediction. In particular, mutual distance constraints consist of two components, the signed distance of each vertex on the human mesh to the scene surface and the distance of basis scene points to the human mesh. We further introduce a global scene representation learned from a signed distance function (SDF) volume to ensure coherence between the global scene representation and the explicit constraint from the mutual distance. We develop a pipeline with two sequential steps: predicting the future mutual …

Stochastic Human Motion Prediction (HMP) aims to predict multiple possible future pose sequences from observed ones. Most prior works learn motion distributions through encoding-decoding in latent space, which does not preserve motion’s spatial-temporal structure. While effective, these methods often require complex, multi-stage training and yield predictions that are inconsistent with the provided history and can be physically unrealistic. To address these issues, we propose CoMusion, a single-stage, end-to-end diffusion-based stochastic HMP framework. CoMusion is inspired from the insight that a smooth future pose initialization improves prediction performance, a strategy not previously utilized in stochastic models but evidenced in deterministic works. To generate such initialization, CoMusion's motion predictor starts with a Transformer-based network for initial reconstruction of corrupted motion. Then, a graph convolutional network (GCN) is employed to refine the prediction considering past observations in the discrete cosine transformation (DCT) space. Our method, facilitated by the Transformer-GCN module design and a proposed variance scheduler, excels in predicting accurate, realistic, and consistent motion, while maintaining an appropriate level of diversity. Experimental results on benchmark datasets demonstrate that CoMusion surpasses prior methods in both accuracy and fidelity, achieving at least a 35% relative improvement in fidelity metrics, while demonstrating superior robustness. Our method, …

Existing 3D human object interaction (HOI) datasets and models simply align global descriptions with the long HOI sequence, while lacking a detailed understanding of intermediate states and the transitions between states. In this paper, we argue that fine-grained semantic alignment, which utilizes state-level descriptions, offers a promising paradigm for learning semantically rich HOI representations. To achieve this, we introduce Semantic-HOI, a new dataset comprising over 20K paired HOI states with fine-grained descriptions for each HOI state and the body movements that happen between two consecutive states. Leveraging the proposed dataset, we design three state-level HOI tasks to accomplish fine-grained semantic alignment within the HOI sequence. Additionally, we propose a unified model called \ModelName, designed to leverage multimodal instructions and empower the Multi-modal Large Language Model to efficiently handle diverse HOI tasks. F-HOI offers multiple advantages: (1) It employs a unified task formulation that supports the use of versatile multimodal inputs. (2) It maintains consistency in HOI across 2D, 3D, and linguistic spaces. (3) It utilizes fine-grained textual supervision for direct optimization, avoiding intricate modeling of HOI states. Extensive experiments reveal that \ModelName effectively aligns HOI states with fine-grained semantic descriptions, adeptly tackling understanding, reasoning, generation, and reconstruction tasks.

The goal of motion understanding is to establish a reliable mapping between motion and action semantics, while it is a challenging many-to-many problem. An abstract action semantic (i.e., walk forwards) could be conveyed by perceptually diverse motions (walking with arms up or swinging), while a motion could carry different semantics w.r.t. its context and intention. This makes an elegant mapping between them difficult. Previous attempts adopted direct-mapping paradigms with limited reliability. Also, current automatic metrics fail to provide reliable assessments of the consistency between motions and action semantics. We identify the source of these problems as the significant gap between the two modalities. To alleviate this gap, we propose Kinematic Phrases (KP) that take the objective kinematic facts of human motion with proper abstraction, interpretability, and generality. Based on KP, we can unify a motion knowledge base and build a motion understanding system. Meanwhile, KP can be automatically converted from motions to text descriptions with no subjective bias, inspiring Kinematic Prompt Generation (KPG) as a novel white-box motion generation benchmark. In extensive experiments, our approach shows superiority over other methods. Our code and data would be made publicly available.

Text-to-motion models excel at efficient human motion generation, but existing approaches lack fine-grained controllability over the generation process. Consequently, modifying subtle postures within a motion or inserting new actions at specific moments remains a challenge, limiting the applicability of these methods in diverse scenarios. In light of these challenges, we introduce CoMo, a Controllable Motion generation model, adept at accurately generating and editing motions by leveraging the knowledge priors of large language models (LLMs). Specifically, CoMo decomposes motions into discrete and semantically meaningful pose codes, with each code encapsulating the semantics of a body part, representing elementary information such as "left knee slightly bent". Given textual inputs, CoMo autoregressively generates sequences of pose codes, which are then decoded into 3D motions. Leveraging pose codes as interpretable representations, an LLM can directly intervene in motion editing by adjusting the pose codes according to editing instructions. Experiments demonstrate that CoMo achieves competitive performance in motion generation compared to state-of-the-art models while, in human studies, CoMo substantially surpasses previous work in motion editing abilities.

Text-to-motion generation requires not only grounding local actions in language but also seamlessly blending these individual actions to synthesize diverse and realistic global motions. However, existing motion generation methods primarily focus on the direct synthesis of global motions while neglecting the importance of generating and controlling local actions. In this paper, we propose the local action-guided motion diffusion model, which facilitates global motion generation by utilizing local actions as fine-grained control signals. Specifically, we provide an automated method for reference local action sampling and leverage graph attention networks to assess the guiding weight of each local action in the overall motion synthesis. During the diffusion process for synthesizing global motion, we calculate the local-action gradient to provide conditional guidance. This local-to-global paradigm reduces the complexity associated with direct global motion generation and promotes motion diversity via sampling diverse actions as conditions. Extensive experiments on two human motion datasets, i.e., HumanML3D and KIT, demonstrate the effectiveness of our method. Furthermore, our method provides flexibility in seamlessly combining various local actions and continuous guiding weight adjustment, accommodating diverse user preferences, which may hold potential significance for the community.

Co-speech gesture video generation is an enabling technique for numerous digital human applications in the post-ChatGPT era. Substantial progress has been made in creating high-quality talking head videos. However, existing hand gesture video generation methods are largely limited by the widely adopted 2D skeleton-based gesture representation, and still struggle to generate realistic hands. We propose a novel end-to-end audio-driven co-speech video generation pipeline to synthesize human speech videos leveraging 3D human mesh-based representations. By adopting a 3D human mesh-based gesture representation, we present a mesh-grounded video generator that includes a mesh texture-map optimization step followed by a new conditional GAN-based network, and outputs photorealistic gesture videos with realistic hands. Our experiments on the TalkSHOW dataset demonstrate the effectiveness of our method over a baseline that uses 2D skeleton-based representation.

We present MOFA-Video, an advanced controllable image animation method that generates video from the given image using various additional controllable signals~(such as human landmarks reference, manual trajectories, and another even provided video) or their combinations. This is different from previous methods which only can work on a specific motion domain or show weak control abilities with diffusion prior. To achieve our goal, we design several domain-aware motion field adapters~(\ie, MOFA-Adapters) to control the generated motions in the video generation pipeline. For MOFA-Adapters, we consider the temporal motion consistency of the video and generate the dense motion flow from the given sparse control conditions first, and then, the multi-scale features of the given image are wrapped as a guided feature for stable video diffusion generation. We naively train two motion adapters for the manual trajectories and the human landmarks individually since they both contain sparse information about the control. After training, the MOFA-Adapters in different domains can also work together for more controllable video generation.

We introduce a novel diffusion-based video generation method, generating a video showing multiple events given multiple individual sentences from the user. Our method does not require a large-scale video dataset since our method uses a pre-trained diffusion-based text-to-video generative model without a fine-tuning process. Specifically, we propose a last frame-aware diffusion process to preserve visual coherence between consecutive videos where each video consists of different events by initializing the latent and simultaneously adjusting noise in the latent to enhance the motion dynamic in a generated video. Furthermore, we find that the iterative update of latent vectors by referring to all the preceding frames maintains the global appearance across the frames in a video clip. To handle dynamic text input for video generation, we utilize a novel prompt generator that transfers course text messages from the user into the multiple optimal prompts for the text-to-video diffusion model. Extensive experiments and user studies show that our proposed method is superior to other video-generative models in terms of temporal coherency of content and semantics.

We present a method to create diffusion-based Video models from pretrained Text-to-Image (T2I) models, overcoming limitations of existing methods. We propose a unique architecture, incorporating a mapping network and frame-wise tokens, tailored for video generation while maintaining the diversity and creativity of the original T2I model. Key innovations include novel loss functions for temporal smoothness and a mitigating gradient sampling technique, ensuring realistic and temporally consistent video generation. Our method, built on the frozen StableDiffusion model, simplifies training processes and allows for seamless integration with off-the-shelf models like ControlNet and DreamBooth. We demonstrate superior performance through extensive experiments and comparisons.

Text-driven video editing has emerged as a prominent application based on the breakthroughs of image diffusion models. Existing state-of-the-art methods focus on zero-shot frameworks due to limited training data and to limit computational expense. To preserve structure consistency, previous frameworks usually employ Denoising Diffusion Implicit Model (DDIM) inversion to provide inverted noise latents as guidance. The key challenge lies in limiting errors caused by the randomness and inaccuracy in each step of the in the naive DDIM inversion process, which can lead to temporal inconsistency in video editing tasks. Our observation indicates that incorporating temporal keyframe information can alleviate the accumulated error during inversion. In this paper, we propose an effective warping strategy in the feature domain to obtain high-quality DDIM inverted noise latents. Specifically, we shuffle the editing frames randomly in each timestep and use optical flow extracted from the source video to propagate the latent features of the first keyframe to subsequent keyframes. Moreover, we develop a comprehensive zero-shot framework that adapts to this strategy in both the inversion and denoising processes, thereby facilitating the generation of consistent edited videos.

Point-drag-based image editing methods, like DragDiffusion, have attracted significant attention. However, point-drag-based approaches suffer from computational overhead and misinterpretation of user intentions, due to the sparsity of point-based editing instructions. In this paper, we propose a region-based copy-and-paste dragging method, RegionDrag, to overcome these limitations. RegionDrag allows users to express their editing instructions in the form of handle and target regions, enabling more precise control and alleviating ambiguity. In addition, region-based operations complete editing in one iteration and are much faster than point-drag-based methods. We also incorporate the attention-swapping technique for enhanced stability during editing. To validate our approach, we extend existing point-drag-based datasets with region-based dragging instructions. Experimental results demonstrate that RegionDrag outperforms existing point-drag-based approaches in terms of speed, accuracy, and alignment with user intentions. Remarkably, RegionDrag completes the edit on an image with a resolution of 512x512 in less than 2 seconds, which is more than 100x faster than DragDiffusion, while achieving better performance. Project page: \url{https://visual-ai.github.io/regiondrag}.

This paper presents a novel approach for real-time image editing leveraging few-shot diffusion models. We demonstrate that disentangled controls can be easily achieved in the few-shot diffusion model by conditioning on a detailed text prompt. Our method involves generating a source image by fixing the random seed and utilizing a lengthy text prompt, followed by modifying one attribute in the text prompt to regenerate the target image. We observe that the source and target images are nearly identical, differing only in the modified attribute. Additionally, we introduce an iterative image inversion technique. The inversion network is conditioned on the input image and the reconstructed image from the previous step, allowing for the correction of the reconstructed image towards the input image. The information of the input image is preserved in the detailed text prompt and four levels of noise maps. To manipulate the inverted image, we freeze the noise maps and modify one attribute in the text prompt, resulting in the generation of a new image similar to the input image with only one attribute changed. Furthermore, our method achieves real-time performance, running in milliseconds for both the inversion and editing processes.

Given a factorization of an image into various components, we present a method to independently control these components through diffusion model sampling. For example, decomposing an image into low and high spatial frequencies allows us to produce images whose low frequencies align with one prompt, and whose high frequencies align with another prompt. That is, we are able to produce hybrid images. We also explore a decomposition into {\it Lab} color space, allowing us to produce images that appear to be one thing when viewed in greyscale, but changes appearance when color is added back. Our method is simple and only modifies the sampling procedure of a pretrained text-conditional image diffusion model. It works by denoising with a composite noise estimate, where each component of the estimate comes from a noise estimate conditioned on a different prompt. We provide qualitative results showing that this method is effective, give intuition for why this approach succeeds, and derive conditions on the image decomposition for the method to work. In addition, we provide quantitative evaluations demonstrating that our method is better than prior work on hybrid image generation, and we generate hybrid images with three different contents.

Handwritten text generation (HTG) conditioned on content and style is a challenging task due to the variability of inter-user characteristics and the unlimited combinations of characters that form new words unseen during training. Diffusion Models have recently shown promising results in HTG, however they are still under-explored. We present DiffusionPen (DiffPen), a 5-shot style handwritten text generation approach based on Latent Diffusion Models. By utilizing a hybrid style extractor that combines the power of metric learning and classification, our approach manages to capture both textual and stylistic characteristics of seen and unseen words and styles and generate realistic handwritten samples. We perform experiments on IAM offline handwriting database to evaluate the generated style and content and compare them with other SotA methods. Our method outperforms other methods qualitatively and quantitatively and additional data from our method can improve the performance of a Handwriting Text Recognition (HTR) system. The code is available at: (code repository will be released in case of acceptance to maintain anonymity).

Methods for finetuning generative models for concept-driven personalization generally achieve strong results for subject-driven or style-driven generation. Recently, low-rank adaptations (LoRA) have been proposed as a parameter-efficient way of achieving concept-driven personalization. While recent work explores the combination of separate LoRAs to achieve joint generation of learned styles and subjects, existing techniques do not reliably address the problem, so that either the subject-fidelity or style-fidelity are compromised. We propose ZipLoRA, a method to cheaply and effectively merge independently trained style and subject LoRAs in order to achieve generation of any user-provided subject in any user-provided style. Experiments on wide range of subject and style combinations show that ZipLoRA can generate compelling results with meaningful improvements over baselines in subject and style fidelity while preserving the ability to recontextualize.

The task of personalized image aesthetic assessment seeks to tailor aesthetic score prediction models to match individual preferences with just a few user-provided inputs. However, the scalability and generalization capabilities of current approaches are considerably restricted by their reliance on an expensive curated database. To overcome this long-standing scalability challenge, we present a unique approach that leverages readily available databases for general image aesthetic assessment and image quality assessment. Specifically, we view each database as a distinct image score regression task that exhibits varying degrees of personalization potential. By determining optimal combinations of task vectors, known to represent specific traits of each database, we successfully create personalized models for individuals. This approach of integrating multiple models allows us to harness a substantial amount of data. Our extensive experiments demonstrate the effectiveness of our approach in generalizing to previously unseen domains---a challenge previous approaches have struggled to achieve---making it highly applicable to real-world scenarios. Our novel approach significantly advances the field by offering scalable solutions for personalized aesthetic assessment and establishing high standards for future research.

Recently, the application of modern diffusion-based text-to-image generation models for creating artistic fonts, traditionally the domain of professional designers, has garnered significant interest. Diverging from the majority of existing studies that concentrate on generating artistic typography, our research aims to tackle a novel and more demanding challenge: the generation of text effects for multilingual fonts. This task essentially requires generating coherent and consistent visual content within the confines of a font-shaped canvas, as opposed to a traditional rectangular canvas. To address this task, we introduce a novel shape-adaptive diffusion model capable of interpreting the given shape and strategically planning pixel distributions within the irregular canvas. To achieve this, we curate a high-quality shape-adaptive image-text dataset and incorporate the segmentation mask as a visual condition to steer the image generation process within the irregular-canvas. This approach enables the traditionally rectangle canvas-based diffusion model to produce the desired concepts in accordance with the provided geometric shapes. Second, to maintain consistency across multiple letters, we also present a training-free, shape-adaptive effect transfer method for transferring textures from a generated reference letter to others. The key insights are building a font effect noise prior and propagating the font effect information in a concatenated latent …

The field of text-to-image (T2I) generation has made significant progress in recent years, thanks to diffusion models. Linguistic control enables effective content creation, but is defective in fine-grained control over image generation. This challenge has been solved in great extent by incorporating additional user-supplied spatial conditions like depth map, edge map into pre-trained T2I model via extra encoding. However, multi-control image synthesis still struggle with input flexibility, handling the relationship among spatial conditions, and maintaining compatibility with text inputs. To address these challenges, we propose AnyControl, a controllable image synthesis framework that supports any combination of various forms of control signals. AnyControl develops a novel multi-control encoder to extract a unified multi-modal embedding for diverse control signals used for guiding the generation process. We achieve this by employing alternate multi-control encoding scheme and multi-control alignment scheme, with learnable queries as a bridge to unite them seamlessly and gradually distill compatible information from spatial conditions guided by textual prompts. This approach enables holistic understanding of user inputs, and produces harmonious results in high quality and fidelity under versatile control signals, demonstrated by extensive quantitative and qualitative results.

Recent breakthroughs in text-to-image diffusion models have significantly advanced the generation of high-fidelity, photo-realistic images from textual descriptions. Yet, these models often struggle with interpreting spatial arrangements from text, hindering their ability to produce images with precise spatial configurations. To bridge this gap, layout-to-image generation has emerged as a promising direction. However, training-based approaches are limited by the need for extensively annotated datasets, leading to high data acquisition costs and a constrained conceptual scope. Conversely, training-free methods face challenges in accurately locating and generating semantically similar objects within complex compositions. This paper introduces a novel training-free approach designed to overcome adversarial semantic intersections during the diffusion conditioning phase. By refining intra-token loss with selective sampling and enhancing the diffusion process with attention redistribution, we propose two innovative constraints: 1) an inter-token constraint that resolves token conflicts to ensure accurate concept synthesis; and 2) a self-attention constraint that improves pixel-to-pixel relationships, enhancing the fidelity and complexity of generated images. Our evaluations confirm the effectiveness of leveraging layout information for guiding the diffusion process, generating content-rich images with enhanced fidelity and complexity.

Diffusion models have become the State-of-the-Art for text-to-image generation, and increasing research effort has been dedicated to adapting the inference process of pretrained diffusion models to achieve zero-shot capabilities. An example is the generation of panorama images, which has been tackled in recent works by combining independent diffusion paths over overlapping latent features, which is referred to as joint diffusion, obtaining perceptually aligned panoramas. However, these methods often yield semantically incoherent outputs and trade-off diversity for uniformity. To overcome this limitation, we propose the Merge-Attend-Diffuse operator, which can be plugged into different types of pretrained diffusion models used in a joint diffusion setting to improve the perceptual and semantical coherence of the generated panorama images. Specifically, we merge the diffusion paths, reprogramming self- and cross-attention to operate on the aggregated latent space. Extensive quantitative and qualitative experimental analysis, together with a user study, demonstrate that our method maintains compatibility with the input prompt and visual quality of the generated images while increasing their semantic coherence.

Text-to-image diffusion models have an unprecedented ability to generate diverse and high-quality images. However, they often struggle to faithfully capture the intended semantics of complex input prompts that include multiple subjects. Recently, numerous layout-to-image extensions have been introduced to improve user control, aiming to localize subjects represented by specific tokens. Yet, these methods often produce semantically inaccurate images, especially when dealing with multiple semantically or visually similar subjects. In this work, we study and analyze the causes of these limitations. Our exploration reveals that the primary issue stems from inadvertent semantic leakage between subjects in the denoising process. This leakage is attributed to the diffusion model’s attention layers, which tend to blend the visual features of different subjects. To address these issues, we introduce Bounded Attention, a training-free method for bounding the information flow in the sampling process. Bounded Attention prevents detrimental leakage among subjects and enables guiding the generation to promote each subject's individuality, even with complex multi-subject conditioning. Through extensive experimentation, we demonstrate that our method empowers the generation of multiple subjects that better align with given prompts and layouts. Our code will be available upon publication.

Personalization is an important topic in text-to-image generation, especially the challenging multi-concept personalization. Current multi-concept methods are struggling with identity preservation, occlusion, and the harmony between foreground and background. In this work, we propose OMG, an occlusion-friendly personalized generation framework designed to seamlessly integrate multiple concepts within a single image. We propose a novel two-stage sampling solution. The first stage takes charge of layout generation and visual comprehension information collection for handling occlusions. The second one utilizes the acquired visual comprehension information and the designed noise blending to integrate multiple concepts while considering occlusions. We also observe that the initial denoising timestep for noise blending is the key to identity preservation and layout. Moreover, our method can be combined with various single-concept models, such as LoRA and InstantID without additional tuning. Especially, LoRA models on civitai.com can be exploited directly. Extensive experiments demonstrate that OMG exhibits superior performance in multi-concept personalization.

The literature on text-to-image generation is plagued by issues of faithfully composing entities with relations. But there lacks a formal understanding of how entity-relation compositions can be effectively learned. Moreover, the underlying phenomenon space that meaningfully reflects the problem structure is not well-defined, leading to an arms race for larger quantities of data in the hope that generalization emerges out of large scale pretraining. We hypothesize that the underlying phenomenological coverage has not been proportionally scaled up, leading to a skew of the presented phenomenon which harms generalization. We introduce statistical metrics that quantify both the linguistic and visual skew of a dataset for relational learning, and show that generalization failures of text-to-image generation is a direct result of an incomplete or unbalanceed phenomenological coverage. We first perform experiments in a synthetic domain and demonstrate that systematically controlled metrics are strongly predictive of generalization performance. Then we move to natural images and show that simple distribution perturbations in light of our theories boost generalization without enlarging the absolute data size. This work informs an important direction towards quality-enhancing the data diversity or balance orthogonal to scaling up the absolute size. Our discussions point out important open questions on 1) Evaluation …

Text-to-image (T2I) generation with Stable Diffusion models (SDMs) involves high computing demands due to billion-scale parameters. To enhance efficiency, recent studies have reduced sampling steps and applied network quantization while retaining the original architectures. The lack of architectural reduction attempts may stem from worries over expensive retraining for such massive models. In this work, we uncover the surprising potential of block pruning and feature distillation for low-cost general-purpose T2I. By removing several residual and attention blocks from the U-Net of SDMs, we achieve 30%~50% reduction in model size, MACs, and latency. We show that distillation retraining is effective even under limited resources: using only 13 A100 days and a tiny dataset, our compact models can imitate the original SDMs (v1.4 and v2.1-base with over 6,000 A100 days). Benefiting from the transferred knowledge, our BK-SDMs deliver competitive results on zero-shot MS-COCO against larger multi-billion parameter models. We further demonstrate the applicability of our lightweight backbones in personalized generation and image-to-image translation. Deployment of our models on edge devices attains 4-second inference. We hope this work can help build small yet powerful diffusion models with feasible training budgets.

Optimizing a text-to-image diffusion model with a given reward function is an important but underexplored research area. In this study, we propose Deep Reward Tuning (DRTune), an algorithm that directly supervises the final output image of a text-to-image diffusion model and back-propagates through the iterative sampling process to the input noise. We find that training earlier steps in the sampling process is crucial for low-level rewards, and deep supervision can be achieved efficiently and effectively by stopping the gradient of the denoising network input. DRTune is extensively evaluated on various reward models. It consistently outperforms other algorithms, particularly for low-level control signals, where all shallow supervision methods fail. Additionally, we fine-tune Stable Diffusion XL 1.0 (SDXL 1.0) model via DRTune to optimize Human Preference Score v2.1, resulting in the Favorable Diffusion XL 1.0 (FDXL 1.0) model. FDXL 1.0 significantly enhances image quality compared to SDXL 1.0 and reaches comparable quality compared with Midjourney v5.2.

Diffusion models, which revolutionized image generation, are facing challenges related to intellectual property. These challenges arise when a generated image is influenced by one or more copyrighted images from the training data. Hence, pinpointing influential images from the training dataset, a task known as data attribution, becomes crucial for the clarity of content origins. We introduce MONTAGE, a pioneering data attribution method. Unlike existing approaches that overlook the internal workings of the training process, MONTAGE integrates a novel technique to monitor generations throughout the training via internal model representations. It is tailored for customized diffusion models, where training access is a practical assumption. This approach, coupled with a new loss function, enables enhanced accuracy as well as granularity of the attributions. The advantage of MONTAGE is evaluated in two granularity levels: Semantic concept (including mix-concept images) and individual image, showing promising results. This underlines MONTAGE's role towards solving copyright concerns in AI-generated digital art and media while enriching the understanding of the generative process.

Text-to-Image (T2I) Diffusion Models (DMs) excel at creating high-quality images from text descriptions but, like many deep learning models, suffer from robustness issues. While there are attempts to evaluate the robustness of T2I DMs as a binary or worst-case problem, they cannot answer how robust in general the model is whenever an adversarial example (AE) can be found. In this study, we first formalise a probabilistic notion of T2I DMs' robustness; and then devise an efficient framework, ProTIP, to evaluate it with statistical guarantees. The main challenges stem from: i) the high computational cost of the image generation process; and ii)identifying if a perturbed input is an AE involves comparing two output distributions, which is fundamentally harder compared to other DL tasks like classification where an AE is identified upon misprediction of labels. To tackle the challenges, we employ sequential analysis with efficacy and futility early stopping rules in the statistical testing for identifying AEs, and adaptive concentration inequalities to dynamically determine the “just-right” number of stochastic perturbations whenever the verification target is met. Empirical experiments validate ProTIP's effectiveness and efficiency, and showcase its application in ranking common defence methods.

Drawing upon StyleGAN's expressivity and disentangled latent space, existing 2D approaches employ textual prompting to edit facial images with different attributes. In contrast, 3D-aware approaches that generate faces at different target poses require attribute-specific classifiers, learning separate model weights for each attribute, and are not scalable for novel attributes. In this work, we propose an efficient, plug-and-play, 3D-aware face editing framework, based on attribute-specific prompt learning, enabling the generation of facial images with controllable attributes across various target poses. To this end, we introduce a text-driven learnable style token-based latent attribute editor (LAE). The LAE harnesses a pre-trained vision-language model to find text-guided attribute-specific editing direction in the latent space of any pre-trained 3D-aware GAN. It utilizes learnable style tokens and style mappers to learn and transform this editing direction to 3D latent space. To train LAE with multiple attributes, we use directional contrastive loss and style token loss. Furthermore, to ensure view consistency and identity preservation across different poses and attributes, we employ several 3D-aware identity and pose preservation losses. Our experiments show that our proposed framework generates high-quality images with 3D awareness and view consistency while maintaining attribute-specific features. We demonstrate the effectiveness of our method on different …

The recent advances in diffusion models (DMs) have revolutionized the generation of realistic and complex images. However, these models also introduce potential safety hazards, such as producing harmful content and infringing data copyrights. Despite the development of safety-driven unlearning techniques to counteract these challenges, doubts about their efficacy persist. To tackle this issue, we introduce an evaluation framework that leverages adversarial prompts to discern the trustworthiness of these safety-driven DMs after they have undergone the process of unlearning harmful concepts. Specifically, we investigated the adversarial robustness of DMs, assessed by adversarial prompts, when eliminating unwanted concepts, styles, and objects. We develop an effective and efficient adversarial prompt generation approach for DMs, termed UnlearnDiffAtk. This method capitalizes on the intrinsic classification abilities of DMs to simplify the creation of adversarial prompts, thereby eliminating the need for auxiliary classification or diffusion models.Through extensive benchmarking, we evaluate the robustness of five widely-used safety-driven unlearned DMs (i.e., DMs after unlearning undesirable concepts, styles, or objects) across a variety of tasks. Our results demonstrate the effectiveness and efficiency merits of UnlearnDiffAtk over the state-of-the-art adversarial prompt generation method and reveal the lack of robustness of current safety-driven unlearning techniques when applied to DMs.

Visual counterfactual explanation (CF) methods modify image concepts, e.g., shape, to change a prediction to a predefined outcome while closely resembling the original query image. Unlike self-explainable models (SEMs) and heatmap techniques, they grant users the ability to examine hypothetical "what-if" scenarios. Previous CF methods either entail post-hoc training, limiting the balance between transparency and CF quality, or demand optimization during inference. To bridge the gap between transparent SEMs and CF methods, we introduce the GdVAE, a self-explainable model based on a conditional variational autoencoder (CVAE), featuring a Gaussian discriminant analysis (GDA) classifier and integrated CF explanations. Full transparency is achieved through a generative classifier that leverages class-specific prototypes for the downstream task and a closed-form solution for CFs in the latent space. The consistency of CFs is improved by regularizing the latent space with the explainer function. Extensive comparisons with existing approaches affirm the effectiveness of our method in producing high-quality CF explanations while preserving transparency.

Recent progress in visual generative models enables the generation of high-quality images. To prevent the misuse of generated images, it is important to identify the origin model that generates them. In this work, we study the origin attribution of generated images in a practical setting where only a few images generated by a source model are available and the source model cannot be accessed. The goal is to check if a given image is generated by the source model. We first formulate this problem as a few-shot one-class classification task. To solve the task, we propose OCC-CLIP, a CLIP-based framework for few-shot one-class classification, enabling the identification of an image's source model, even among multiple candidates. Extensive experiments corresponding to various generative models verify the effectiveness of our OCC-CLIP framework. Furthermore, an experiment based on the recently released DALL·E-3 API verifies the real-world applicability of our solution.

Latent Diffusion Models (LDMs) are powerful and potential tools for facilitating generation-based methods for domain generalization. However, existing diffusion-based DG methods are restricted to offline augmentation using LDM and suffer from degraded performance and prohibitive computational costs. To address these challenges, we propose DomainFusion to simultaneously achieve knowledge extraction in the latent space and augmentation in the pixel space of the Latent Diffusion Model (LDM) for efficiently and sufficiently exploiting LDM. We develop a Latent Distillation module that distills gradient priors from LDM to guide the optimization of DG models. Moreover, we design an online lightweight augmentation method by decomposing candidate images into styles and contents for using LDM in a fast and online fashion. Experimental results demonstrate that DomainFusion outperforms diffusion-based methods by a large margin and achieves SOTA performance on existing DG benchmark datasets. Remarkably, DomainFusion can significantly reduce the number of generated images (e.g. by more than 97% on DomainNet) without finetuning LDM.

Text-to-image diffusion models have shown remarkable success in synthesizing photo-realistic images. Apart from creative applications, can we use such models to synthesize samples that aid the few-shot training of discriminative models? In this work, we propose AlignDiff, a general framework for synthesizing training images and masks for few-shot segmentation. We identify two crucial misalignments that arise when utilizing pre-trained diffusion models in segmentation tasks, which need to be addressed to create realistic training samples and align the synthetic data distribution with the real training distribution: 1) instance-level misalignment, where generated samples of rare categories are often misaligned with target tasks) and 2) annotation-level misalignment, where diffusion models are limited to generating images without pixel-level annotations. AlignDiff overcomes both challenges by leveraging a few real samples to guide the generation, thus improving novel IoU over baseline methods in few-shot segmentation and generalized few-shot segmentation on Pascal-5i and COCO-20i by up to 80%. Notably, AlignDiff is capable of augmenting the learning of out-of-distribution uncommon categories on FSS-1000, while naive diffusion model generates samples that diminish segmentation performance.

The rise of billion-parameter diffusion models such as Stable Diffusion XL, Imagen, and Dall-E3 significantly propels the domain of generative AI. However, their large-scale architecture presents challenges in fine-tuning and deployment due to high resource demands and slow inference speed. This paper delves into relatively unexplored yet promising realm of fine-tuning quantized diffusion models. Our analysis identified that the baseline neglects the distinct pattern in model weights and different roles throughout time-step when finetuning the diffusion model. To address these limitations, we introduce a novel memory-efficient fine-tuning framework directly applicable to quantized diffusion models, dubbed TuneQDM. Our approach introduces quantization scales as separable functions to consider inter-channel patterns of weight and optimizes scales in a time-step specific manner for effective reflection of the role of time-step. TuneQDM demonstrates performance on par with its full-precision counterpart, while simultaneously offering a substantial advantage in terms of memory efficiency. The experimental results demonstrate that our efficient framework consistently outperforms the baseline in single-/multi-subject generation, exhibiting high subject fidelity and prompt fidelity comparable to the full precision model.

Diffusion models have become a mainstream approach for high-resolution image synthesis. However, directly generating \textbf{higher-resolution} images from pretrained diffusion models will encounter unreasonable object duplication and exponentially increase the generation time. In this paper, we discover that object duplication arises from feature duplication in the deep blocks of the U-Net. Concurrently, We pinpoint the extended generation times to self-attention redundancy in U-Net's top blocks. To address these issues, we propose a tuning-free higher-resolution framework named HiDiffusion. Specifically, HiDiffusion contains Resolution-Aware U-Net~(RAU-Net) that dynamically adjusts the feature map size to resolve object duplication and engages Modified Shifted Window Multi-head Self-Attention~(MSW-MSA) that utilizes optimized window attention to reduce computations. we can integrate HiDiffusion into various pretrained diffusion models to scale image generation resolutions even to 4096×4096 at 1.5-6× the inference speed of previous methods. Extensive experiments demonstrate that our approach can address object duplication and heavy computation issues, achieving state-of-the-art performance on higher-resolution image synthesis tasks.

We introduce EGIC, an enhanced generative image compression method that allows traversing the distortion-perception curve efficiently from a single model. EGIC is based on two novel building blocks: i) OASIS-C, a conditional pre-trained semantic segmentation-guided discriminator, which provides both spatially and semantically-aware gradient feedback to the generator, conditioned on the latent image distribution, and ii) Output Residual Prediction (ORP), a retrofit solution for multi-realism image compression that allows control over the synthesis process by adjusting the impact of the residual between an MSE-optimized and GAN-optimized decoder output on the GAN-based reconstruction. Together, EGIC forms a powerful codec, outperforming state-of-the-art diffusion and GAN-based methods (e.g., HiFiC, MS-ILLM, and DIRAC-100), while performing almost on par with VTM-20.0 on the distortion end. EGIC is simple to implement, very lightweight, and provides excellent interpolation characteristics, which makes it a promising candidate for practical applications targeting the low bit range.

Existing natural image matting algorithms inevitably have flaws in their predictions on difficult cases, and their one-step prediction manner cannot further correct these errors. In this paper, we investigate a multi-step iterative approach for the first time to tackle the challenging natural image matting task, and achieve excellent performance by introducing a pixel-level denoising diffusion method (DiffMatte) for the alpha matte refinement. To improve iteration efficiency, we design a lightweight diffusion decoder as the only iterative component to directly denoise the alpha matte, saving the huge computational overhead of repeatedly encoding matting features. We also propose an ameliorated self-aligned strategy to consolidate the performance gains brought about by the iterative diffusion process. This allows the model to adapt to various types of errors by aligning the noisy samples used in training and inference, mitigating performance degradation caused by sampling drift. Extensive experimental results demonstrate that DiffMatte not only reaches the state-of-the-art level on the mainstream Composition-1k test set, surpassing the previous best methods by 8% and 15 in the SAD metric and MSE metric respectively, but also show stronger generalization ability in other benchmarks. The code will be open-sourced for the following research and applications.

Diffusion models have achieved remarkable success across a range of generative tasks. Recent efforts to enhance diffusion model architectures have reimagined them as a form of multi-task learning, where each task corresponds to a denoising task at a specific noise level. While these efforts have focused on parameter isolation and task routing, they fall short of capturing detailed inter-task relationships and risk losing semantic information, respectively. In response, we introduce Switch Diffusion Transformer (Switch-DiT), which establishes inter-task relationships between conflicting tasks without compromising semantic information. To achieve this, we employ a sparse mixture-of-experts within each transformer block to utilize semantic information and facilitate handling conflicts in tasks through parameter isolation. Additionally, we propose a diffusion prior loss, encouraging similar tasks to share their denoising paths while isolating conflicting ones. Through these, each transformer block contains a shared expert across all tasks, where the common and task-specific denoising paths enable the diffusion model to construct its beneficial way of synergizing denoising tasks. Extensive experiments validate the effectiveness of our approach in improving both image quality and convergence rate, and further analysis demonstrates that Switch-DiT constructs tailored denoising paths across various generation scenarios.

We present MoE-DiffIR, an innovative universal compressed image restoration (CIR) method with task-customized diffusion priors. This intends to handle two pivotal challenges in the existing CIR methods: (i) lacking adaptability and universality for different image codecs, \eg, JPEG and WebP; (ii) poor texture generation capability, particularly at low bitrates. Specifically, our MoE-DiffIR develops the powerful mixture-of-experts (MoE) prompt module, where some basic prompts cooperate to excavate the task-customized diffusion priors from Stable Diffusion (SD) for each compression task. Moreover, the degradation-aware routing mechanism is proposed to enable the flexible assignment of basic prompts. To activate and reuse the cross-modality generation prior of SD, we design the visual-to-text adapter for MoE-DiffIR, which aims to adapt the embedding of low-quality images from the visual domain to the textual domain as the textual guidance for SD, enabling more consistent and reasonable texture generation. We also construct one comprehensive benchmark dataset for universal CIR, covering 21 types of degradations from 7 popular traditional and learned codecs. Extensive experiments on universal CIR have demonstrated the excellent robustness and texture restoration capability of our proposed MoE-DiffIR.

Neural networks trained end-to-end give state-of-the-art performance for image denoising. However, when applied to an image outside of the training distribution, the performance often degrades significantly. In this work, we propose a test-time training (TTT) method based on masked image modeling (MIM) to improve denoising performance for out-of-distribution images. The method, termed TTT-MIM, consists of a training stage and a test time adaptation stage. At training, we minimize a standard supervised loss and a self-supervised loss aimed at reconstructing masked image patches. At test-time, we minimize a self-supervised loss to fine-tune the network to adapt to a single noisy image. Experiments show that our method can improve performance under natural distribution shifts, in particular it adapts well to real-world camera and microscope noise. A competitor to our method of training and finetuning is to use a zero-shot denoiser that does not rely on training data. However, compared to state-of-the-art zero-shot denoisers, our method shows superior performance, and is much faster, suggesting that training and finetuning on the test instance is a more efficient approach to image denoising than zero-shot methods in setups where little to no data is available.

All-in-one image restoration aims to handle multiple degradation types using one model. We propose a simple pipeline for all-in-one blind image restoration to Restore Anything with Masks (RAM). We focus on the image content itself by utilizing the MIM to extract intrinsic image information rather than distinguishing degradation types like other methods. Our pipeline consists of two stages: masked image pre-training and fine-tuning with mask attribute conductance. We design a simple masking pre-training approach tailored to all-in-one image restoration, boosting networks to focus more on extracting image content priors from any degradation, which turns out to be a more balanced (between different restoration tasks) and stronger performance. To bridge the gap of input integrity while preserving learned image priors as much as possible, we selectively fine-tuned a small portion of the layers. Specifically, the importance of each layer is ranked by the proposed Mask Attribute Conductance (MAC), and the layers with higher contributions are selected for finetuning. Extensive quantitative and qualitative experiments demonstrate that our method achieves state-of-the-art performance. Our code and model will be released.

Real-world image super-resolution deals with complex and unknown degradations, making it challenging to produce plausible results in a single step. In this work, we propose a transformer model with an iterative generation process that iteratively refines the results based on predicted confidences. It allows the model to focus on regions with low confidences and generate more confident and accurate results. Specifically, our model learns to predict the visual tokens of the high-resolution image and their corresponding confidence scores, conditioned on the low-resolution image. By keeping only the most confident tokens at each iteration and re-predicting the other tokens in the next iteration, our model generates all high-resolution tokens within a few steps. To ensure consistency with the low-resolution input image, we further propose a conditional controlling module that utilizes the low-resolution image to control the decoding process from high-resolution tokens to image pixels. Experiments demonstrate that our model achieves state-of-the-art performance on real-world datasets while requiring fewer iteration steps compared to recent diffusion models.

Most current single image-deraining (SID) methods are based on the Transformer, which brings global modeling capabilities and is critical for high-quality reconstruction. However, their architectures only consider constructing long-range dependencies from the spatial domain, which suffers from a significant computational burden to keep effectiveness. Besides, these methods either overlook negative sample information in the optimization pipeline or underutilize the rain streak characteristics present in the negative ones. To tackle these problems, we propose a Frequency-Aware Deraining Transformer Framework (FADformer) that fully captures frequency domain features to achieve efficient rain removal. Specifically, we first construct the FADBlock, which comprises the Fused Fourier Convolution Mixer (FFCM) and Prior-Gated Feed-forward Network (PGFN). Unlike self-attention mechanisms, the FFCM exclusively conducts convolution operations in both spatial and frequency domains, endowing it with local-global capturing capabilities and efficiency. Simultaneously, the PGFN introduces residue channel prior in a gate-controlled manner to both enhance local details and retain the structure of features. Furthermore, we introduce a Frequency-domain Contrastive Regularization (FCR) during the training phase. The FCR facilitates contrastive learning in the frequency domain, enhancing the contribution of rain streak patterns in negative samples to improve performance. Extensive experiments on synthetic and real-world datasets show that the proposed method …

Blind image deconvolution (BID) is a classic yet challenging problem in the field of image processing. Recent advances in deep image prior (DIP) have motivated a series of DIP-based approaches, demonstrating remarkable success in BID, particularly in scenarios with motion blurring. However, due to the high non-convexity of the inherent optimization process, these methods are notorious for their sensitivity to the initialized kernel. To alleviate this issue and further improve their performance, we propose a new framework for BID that better considers the prior modeling and the initialization for blur kernels, in particular of motion blur, leveraging a deep generative model. The proposed approach pre-trains a generative adversarial network-based kernel generator that aptly characterizes the kernel priors and a kernel initializer that facilitates a well-informed initialization for the blur kernel through latent space encoding. With the pre-trained kernel generator and initializer, one can obtain a high-quality initialization of the blur kernel, and enable optimization within a compact latent kernel manifold. Such a framework results in an evident performance improvement over existing DIP-based BID methods. Extensive experiments on different datasets demonstrate the effectiveness of the proposed method. Notably, it achieves state-of-the-art performance on the challenging benchmark by Lai et al.

Reconstructing High Dynamic Range (HDR) image from multiple Low Dynamic Range (LDR) images with different exposures is a challenging task when facing truncated texture and complex motion. Existing deep learning-based methods have achieved great success by either following the alignment and fusion pipeline or utilizing attention mechanism. However, the large computation cost and inference delay hinder them from deploying on resource limited devices. In this paper, to achieve better efficiency, a novel Selective Alignment Fusion Network (SAFNet) for HDR imaging is proposed. After extracting pyramid features, it jointly refines valuable area masks and cross-exposure motion in selected regions with shared decoders, and then fuses high quality HDR image in an explicit way. This approach can focus the model on finding valuable regions while estimating their easily detectable and meaningful motion. For further detail enhancement, a lightweight refine module is introduced which enjoys privileges from previous optical flow, selection masks and initial prediction. Moreover, to facilitate learning on samples with large motion, a new window partition cropping method is presented during training. Experiments on public and newly developed challenging datasets show that proposed SAFNet not only exceeds previous SOTA competitors quantitatively and qualitatively, but also runs order of magnitude faster. Code …

In this work, we investigate the understudied effect of the training data used for image super-resolution (SR). Most commonly, novel SR methods are developed and benchmarked on common training datasets such as DIV2K and DF2K. However, we investigate and rethink the training data from the perspectives of diversity and quality, thereby addressing the question of "How important is SR training for SR models?". To this end, we propose an automated image evaluation pipeline. With this, we stratify existing high-resolution image datasets and larger-scale image datasets such as ImageNet and PASS to compare their performances. We find that datasets with (i) low compression artifacts, (ii) high within-image diversity as judged by the number of different objects, and (iii) a large number of images from ImageNet or PASS all positively affect SR performance. We hope that the proposed simple-yet-effective dataset curation pipeline will inform the construction of SR datasets in the future and yield overall better models.

Single Image Super-Resolution (SISR) plays a vital role in various applications, driven by advancements in Deep Neural Networks (DNNs). However, increasing model complexity raises computational costs, necessitating efficient solutions. Existing patch-based approaches aiming at efficient SR encounter challenges in adapting to varying pixel difficulties and suffer from decreased efficiency with larger patches. To address these limitations, we propose Pixel-level Classifier for Single Image Super-Resolution (PCSR), a novel pixel-level distribution method for efficient SISR. Our approach optimizes computational resource allocation at the pixel level, achieving better efficiency compared to patch-based methods, and also provides user tunability without re-training and mitigates artifacts through post-processing techniques. Experimental results demonstrate the effectiveness of PCSR across diverse SISR models and benchmarks, surpassing existing approaches in terms of the PSNR-FLOPs trade-off.

Quantization is a promising approach to reduce the high computational complexity of image super-resolution (SR) networks. However, low-bit quantization leads to severe accuracy loss in SR networks compared to high-level tasks such as image classification. This is because the feature distributions of the SR networks are significantly divergent for each channel or input image, making it difficult to determine a quantization range. Existing SR quantization works approach this distribution mismatch problem by dynamically adapting quantization ranges to the variant distributions during the test time. However, such a dynamic adaptation incurs additional computational costs that limit the benefits of quantization. Instead, we propose a new quantization-aware training framework that effectively overcomes the distribution mismatch problem in SR networks without the need for dynamic adaptation. Intuitively, the mismatch can be reduced by directly regularizing the distance between the feature to be quantized and the quantization grids during training. However, we observe that mismatch regularization can collide with reconstruction loss during training and adversely affect the SR accuracy. Thus, we avoid the conflict between two losses by regularizing the mismatch only when the gradients of mismatch regularization are cooperative with those of reconstruction loss. Additionally, we introduce a layer-wise weight clipping correcting scheme …

With the rapid advancement of stereo vision technologies, stereo image compression has emerged as a crucial field that continues to draw significant attention. Previous approaches have primarily employed a unidirectional paradigm, where the compression of one view is dependent on the other, resulting in imbalanced compression. To address this issue, we introduce a symmetric bidirectional stereo image compression architecture, named BiSIC. Specifically, we propose a 3D convolution based codec backbone to capture local features and incorporate bidirectional attention blocks to exploit global features. Moreover, we design a novel cross-dimensional entropy model that integrates various conditioning factors, including the spatial context, channel context, and stereo dependency, to effectively estimate the distribution of latent representations for entropy coding. Extensive experiments demonstrate that our proposed BiSIC outperforms conventional image/video compression standards, as well as state-of-the-art learning-based methods, in terms of both PSNR and MS-SSIM.

Recently, learning-based Hyperspectral image (HSI) reconstruction methods have demonstrated promising performance and dominated the mainstream research direction. However, existing learning-based methods still have two issues. 1) Unable to consider both the spatial sparsity and the inter-spectral similarity prior of HSI. 2) Treat all regions equally, ignoring that texture-rich regions and edge regions are more difficult to reconstruct than smooth regions. To address these issues, we propose an uncertainty-driven HSI reconstruction method termed Specformer. Specifically, we first introduce a frequency-wise self-attention (FWSA) and combine it with spatial-wise local-window self-attention (LWSA) with a parallel design to form a Spatial-Frequency (SF) block. LWSA can guide the network to focus on the regions with dense spectral information, and FWSA can capture the inter-spectral similarity. Parallel design helps the network to model cross-window connections, expand its receptive fields while maintaining linear complexity. We use SF-block as the main building block in a multi-scale U-shape network to form our Specformer. In addition, we introduce an uncertainty-driven self-adaptive loss function, which can reinforce the network's attention to challenging regions with rich textures and edges. Comprehensive experiments show that our Specformer significantly outperforms state-of-the-art methods on simulation and real HSI datasets while requiring cheaper computational and memory costs. …

Compressed sensing (CS) has emerged to overcome the inefficiency of Nyquist sampling. However, traditional optimization-based reconstruction is slow and can not yield an exact image in practice. Deep learning-based reconstruction has been a promising alternative to optimization-based reconstruction, outperforming it in accuracy and computation speed. Finding an efficient sampling method with deep learning-based reconstruction, especially for Fourier CS remains a challenge. Existing joint optimization of sampling-reconstruction works (H1) optimize the sampling mask but have low potential as it is not adaptive to each data point. Adaptive sampling (H2) has also disadvantages of difficult optimization and Pareto sub-optimality. Here, we propose a novel adaptive selection of sampling-reconstruction (H1.5) framework that selects the best sampling mask and reconstruction network for each input data. We provide theorems that our method has a higher potential than H1 and effectively solves the Pareto sub-optimality problem in sampling-reconstruction by using separate reconstruction networks for different sampling masks. To select the best sampling mask, we propose to quantify the high-frequency Bayesian uncertainty of the input, using a super-resolution space generation model. Our method outperforms joint optimization of sampling-reconstruction (H1) and adaptive sampling (H2) by achieving significant improvements on several Fourier CS problems.

Image reconstruction from incomplete measurements is a basic task in medical imaging. While supervised deep learning proves to be a powerful tool for image reconstruction, it demands a substantial number of latent images for training. To extend the application of deep learning to medical imaging where collecting latent images poses challenges, this paper introduces an unsupervised test-time adaptation approach. The proposed approach leverages a pre-trained model on an external dataset and efficiently adapts it to each test sample so that the model performs optimally on each specific sample. Model adaption is done by introducing an unrolling network with additional lightweight adaptive linear convolution layers, enabling efficient alignment of testing samples with the distribution targeted in the pre-trained model. This approach is inspired by the connection between linear convolutional network and Wiener filtering. Extensive experiments showed significant performance gain of the proposed method over other unsupervised methods and model adaptation techniques in two medical imaging tasks.

Biomedical imaging datasets are often small and biased, meaning that real-world performance of predictive models can be substantially lower than expected from internal testing. This work proposes using generative image editing to simulate dataset shifts and diagnose failure modes of biomedical vision models; this can be used in advance of deployment to assess readiness, potentially reducing cost and patient harm. Existing editing methods can produce undesirable changes, with spurious correlations learned due to the co-occurrence of disease and treatment interventions, limiting practical applicability. To address this, we train a text-to-image diffusion model on multiple chest X-ray datasets and introduce a new editing method, RadEdit, that uses multiple image masks, if present, to constrain changes and ensure consistency in the edited images, minimising bias. We consider three types of dataset shifts: acquisition shift, manifestation shift, and population shift, and demonstrate that our approach can diagnose failures and quantify model robustness without additional data collection, complementing more qualitative tools for explainable AI.

Deep neural networks (DNNs) are frequently employed in a variety of computer vision applications. Nowadays, an emerging trend in the current video distribution system is to take advantage of DNN's overfitting properties to perform video resolution upscaling. By splitting videos into chunks and applying a super-resolution (SR) model to overfit each chunk, this scheme of SR models plus video chunks is able to replace traditional video transmission to enhance video quality and transmission efficiency. However, many models and chunks are needed to guarantee high performance, which leads to tremendous overhead on model switching and memory footprints at the user end. To resolve such problems, we propose a \underline{Dy}namic \underline{D}eep neural network assisted by a \underline{C}ontent-\underline{A}ware data processing pipeline to reduce the model number down to one (Dy-DCA), which helps promote performance while conserving computational resources. Additionally, to achieve real acceleration on the user end, we designed a framework that optimizes dynamic features (e.g., dynamic shapes, sizes, and control flow) in Dy-DCA to enable a series of compilation optimizations, including fused code generation, static execution planning, etc. By employing such techniques, our method achieves better PSNR and real-time performance (33 FPS) on an off-the-shelf mobile phone. Meanwhile, assisted by our compilation …

Despite the abundant availability and content richness for video data, its high-dimensionality poses challenges for video research. Recent advancements have explored the implicit representation for videos using deep neural networks, demonstrating strong performance in applications such as video compression and enhancement. However, the prolonged encoding time remains a persistent challenge for video Implicit Neural Representations (INRs). In this paper, we focus on improving the speed of video encoding and decoding within implicit representations. We introduce two key components: NeRV-Enc, a transformer-based hyper-network for fast encoding; and NeRV-Dec, an efficient video loader designed to streamline video research. NeRV-Enc achieves an impressive speed-up of by eliminating gradient-based optimization. Meanwhile, NeRV-Dec simplifies video decoding, outperforming conventional codecs with a loading speed faster, and surpassing RAM loading with pre-decoded videos ( faster while being smaller in size).

Cross-modal steganography is committed to hiding secret information of one modality in another modality. Despite the advancement in the field of steganography by the introduction of deep learning, cross-modal steganography still remains to be a challenge to the field. The incompatibility between different modalities not only complicate the hiding process but also results in increased vulnerability to detection. To rectify these limitations, we present INRSteg, an innovative cross-modal steganography framework based on Implicit Neural Representations (INRs). We introduce a novel network allocating framework with a masked parameter update which facilitates hiding multiple data and enables cross modality across image, audio, video and 3D shape. Moreover, we eliminate the necessity of training a deep neural network and therefore substantially reduce the memory and computational cost and avoid domain adaptation issues. To our knowledge, in the field of steganography, this is the first to introduce diverse modalities to both the secret and cover data. Detailed experiments in extreme modality settings demonstrate the flexibility, security, and robustness of INRSteg.

Generative AI raises many societal concerns such as boosting disinformation and propaganda campaigns. Watermarking AI-generated content is a key technology to address these concerns and has been widely deployed in industry. However, watermarking is vulnerable to removal attacks and forgery attacks. In this work, we propose the first image watermarks with certified robustness guarantees against removal and forgery attacks. Our method leverages randomized smoothing, a popular technique to build certifiably robust classifiers and regression models. Our major technical contributions include extending randomized smoothing to watermarking by considering its unique characteristics, deriving the certified robustness guarantees, and designing algorithms to estimate them. Moreover, we extensively evaluate our image watermarks in terms of both certified and empirical robustness.

Early smoke segmentation (ESS) plays a crucial role in accurately locating the source of smoke, facilitating prompt fire rescue operations and gas leak detection. Unlike regular objects, which are typically rigid, opaque, and have clear boundaries, ESS presents challenges due to the large areas of high transparency in early smoke. This leads to a significant similarity between smoke features and the surrounding background features. The key solution is to obtain a discriminative embedding space. Some distance-based methods have pursued this goal by using specific loss functions (e.g., pair-based Triplet loss and proxy-based NCA loss) to constrain the feature extractor. In this paper, we propose a novel approach called discriminative scatter analysis (DSA). Instead of solely measuring Euclidean distance, DSA assesses the compactness and separation of the embedding space from a sample scatter perspective. DSA is performed on both pixel-proxy scatter (IOS) and proxy-proxy scatter (OOS), and a unified loss function is designed to optimize the feature extractor. DSA can be easily integrated with regular segmentation methods. It is applied only during training and without incurring any additional computational cost during inference. Extensive experiments have demonstrated that DSA can consistently improve the performance of various models in ESS.

The rapid development of image processing and manipulation techniques poses unprecedented challenges in multimedia forensics, especially in Image Forgery Localization (IFL). This paper addresses two key challenges in IFL: (1) Various forgery techniques leave distinct forensic traces. However, existing models overlook variations among forgery patterns. The diversity of forgery techniques makes it challenging for a single static detection method and network structure to be universally applicable. To address this, we propose AdaIFL, a dynamic IFL framework that customizes various expert groups for different network components, constructing multiple distinct feature subspaces. By leveraging adaptively activated experts, AdaIFL can capture discriminative features associated with forgery patterns, thereby enhancing the model's generalization ability. (2) Many forensic traces and artifacts are located at the boundaries of the forged region. Existing models either ignore the differences in discriminative information or use edge supervision loss to force the model to focus on the region boundaries. This hard-constrained approach is prone to attention bias, causing the model to be overly sensitive to image edges or fail to finely capture all forensic traces. To address this, we propose a feature importance-aware attention, a flexible approach that adaptively perceives the importance of different regions and aggregates region features into …

Face anti-spoofing (FAS) plays a vital role in preventing face recognition (FR) systems from presentation attacks. Nowadays, FAS systems face the challenge of domain shift, impacting the generalization performance of existing FAS methods. In this paper, we rethink about the inherence of domain shift and deconstruct it into two factors: image style and image quality. Quality influences the purity of the presentation of spoof information, while style affects the manner in which spoof information is presented. Based on our analysis, we propose DiffFAS framework, which quantifies quality as prior information input into the network to counter image quality shift, and performs diffusion-based high-fidelity cross-domain and cross-attack types generation to counter image style shift. DiffFAS transforms easily collectible live faces into high-fidelity attack faces with precise labels while maintaining consistency between live and spoof face identities, which can also alleviate the scarcity of labeled data with novel type attacks faced by nowadays FAS system. We demonstrate the effectiveness of our framework on challenging cross-domain and cross-attack FAS datasets, achieving the state-of-the-art performance. The code shall be released on the GitHub.

Understanding the vulnerability of face recognition system to malicious attacks is of critical importance. Previous works have focused on reconstructing face images that can penetrate a targeted verification system. Even in the white-box scenario, however, naively reconstructed images misrepresent the identity information, hence the attacks are easily neutralized once the face system is updated or changed. In this paper, we aim to reconstruct face images which are capable of transferring face attacks on unseen networks. We term this problem as Face Reconstruction Transfer Attack (FRTA) and show that it can be formulated as an out-of-distribution (OOD) generalization problem. Inspired by its OOD nature, we propose to solve FRTA by Averaged Latent Search and Unsupervised Validation with pseudo target (ALSUV). To strengthen the reconstruction attack on OOD unseen networks, ALSUV reconstructs the face by searching the latent of amortized generator StyleGAN2 through multiple latent optimization, latent optimization trajectory averaging, and unsupervised validation with a pseudo target. We demonstrate the efficacy and generalization of our method on widely used face datasets, accompanying it with extensive ablation studies and analyses visually, qualitatively, and quantitatively. The source code will be released.

Facial affective behavior analysis (FABA) is crucial for understanding human mental states from images. However, traditional approaches primarily deploy models to discriminate among discrete emotion categories, and lack the fine granularity and reasoning capability for complex facial behaviors. The advent of Multi-modal Large Language Models (MLLMs) has been proven successful in general visual understanding tasks. However, directly harnessing MLLMs for FABA is challenging due to the scarcity of datasets and benchmarks, neglecting facial prior knowledge, and low training efficiency. To address these challenges, we introduce (i) an instruction-following dataset for two FABA tasks, e.g., emotion and action unit recognition, (ii) a benchmark FABA-Bench with a new metric considering both recognition and generation ability, and (iii) a new MLLM ''EmoLA'' as a strong baseline to the community. Our initiative on the dataset and benchmarks reveal the nature and rationale of facial affective behaviors, i.e., fine-grained facial movement, interpretability, and reasoning. Moreover, to build an effective and efficient FABA MLLM, we introduce a facial prior expert module with face structure knowledge and a low-rank adaptation module into pre-trained MLLM. We conduct extensive experiments on FABA-Bench and four commonly-used FABA datasets. The results demonstrate that the proposed facial prior expert can boost the …

With the rise of digital media content production, the need for analyzing movies and TV series episodes to locate the main cast of characters precisely is gaining importance. Specifically, Video Face Clustering aims to group together detected video face tracks with common facial identities. This problem is very challenging due to the large range of pose, expression, appearance, and lighting variations of a given face across video frames. Generic pre-trained Face Identification (ID) models fail to adapt well to the video production domain, given its high dynamic range content and also unique cinematic style. Furthermore, traditional clustering algorithms depend on hyperparameters requiring individual tuning across datasets. In this paper, we present a novel video face clustering approach that learns to adapt a generic face ID model to new video face tracks in a fully self-supervised fashion. We also propose a parameter-free clustering algorithm that is capable of automatically adapting to the finetuned model's embedding space for any input video. Due to the lack of comprehensive movie face clustering benchmarks, we also present a first-of-kind movie dataset: MovieFaceCluster. Our dataset is handpicked by film industry professionals and contains extremely challenging face ID scenarios. Experiments show our method's effectiveness in handling difficult …

Scaling up the size of vision models has been the de facto standard to obtain more powerful visual representations. In this work, we discuss the point beyond which larger vision models are not necessary. First, we demonstrate the power of Scaling on Scales (S^2), whereby a pre-trained and frozen smaller vision model (e.g., ViT-B or ViT-L), run over multiple image scales, can outperform larger models (e.g., ViT-H or ViT-G) on classification, segmentation, depth estimation, Multimodal LLM (MLLM) benchmarks, and robotic manipulation. Notably, S^2 achieves state-of-the-art performance in detailed understanding of MLLM on V* benchmark, surpassing models such as GPT-4V. We examine the conditions under which S^2 is a preferred scaling approach compared to scaling on model size. While larger models have the advantage of better generalization on hard examples, we show that features of larger vision models can be well approximated by those of multi-scale smaller models. This suggests most, if not all, of the representations learned by current large pre-trained models can also be obtained from multi-scale smaller models. Our results confirm that a multi-scale smaller model has comparable learning capacity to a larger model, and show that pre-training smaller models with S^2 can match or even exceed the …

Panoramic images, capturing a 360° field of view (FoV), encompass omnidirectional spatial information crucial for scene understanding. However, it is not only costly to obtain training-sufficient dense-annotated panoramas but also application-restricted when training models in a close-vocabulary setting. To tackle this problem, in this work, we define a new task termed Open Panoramic Segmentation (OPS), where models are trained with FoV-restricted pinhole images with labeled data in an open-vocabulary setting while evaluated with FoV-open panoramic images, enabling the zero-shot open panoramic segmentation ability of models. Moreover, we propose a model named OOOPS with a Deformable Adapter Network (DAN), which significantly improves zero-shot panoramic semantic segmentation performance. To further enhance the distortion-aware modeling ability from the pinhole source domain, we propose a novel data augmentation method called Random Equirectangular Projection (RERP) which is specifically designed to address object deformations in advance. Surpassing other state-of-the-art open-vocabulary semantic segmentation approaches, a remarkable performance boost on three panoramic datasets, WildPASS, Stanford2D3D, and Matterport3D, proves the effectiveness of our proposed OOOPS model with RERP on the OPS task, especially +2.2% on outdoor WildPASS and +2.4% mIoU on indoor Stanford2D3D. The source code will be made publicly available.

Tracking multiple objects in a continuous video stream is crucial for many computer vision tasks. It involves detecting and associating objects with their respective identities across successive frames. Despite significant progress made in multiple object tracking (MOT), recent studies have revealed the vulnerability of existing MOT methods to adversarial attacks. Nevertheless, all of these attacks belong to digital attacks that inject pixel-level noise into input images, and are therefore ineffective in physical scenarios. To fill this gap, we propose PapMOT, which can generate physical adversarial patches against MOT for both digital and physical scenarios. Besides attacking the detection mechanism, PapMOT also optimizes a printable patch that can be detected as new targets to mislead the identity association process. Moreover, we introduce a patch enhancement strategy to further degrade the temporal consistency of tracking results across video frames, resulting in more aggressive attacks. We further develop new evaluation metrics to assess the robustness of MOT against such attacks. Extensive evaluations on multiple datasets demonstrate that our PapMOT can successfully attack various architectures of MOT trackers in digital scenarios. We also validate the effectiveness of PapMOT for physical attacks by deploying printed adversarial patches in the real world.

We present a simple, self-supervised approach to the Tracking Any Point (TAP) problem. We train a global matching transformer to find cycle consistent tracks through video via contrastive random walks, using the transformer's attention-based global matching to define the transition matrices for a random walk on a space-time graph. The ability to perform "all pairs" comparisons between points allows the model to obtain high spatial precision and to obtain a strong contrastive learning signal, while avoiding the complexities of recent approaches (such as coarse-to-fine matching). To do this, we propose a number of design decisions that allow global matching architectures to be trained through self-supervision using cycle consistency. For example, we identify that transformer-based methods are sensitive to shortcut solutions, and propose a data augmentation scheme to address them. Our method achieves strong performance on the TapVid benchmarks, outperforming previous self-supervised tracking methods, such as DIFT, and is competitive with several supervised methods.

WiFi-based human sensing has exhibited remarkable potential to analyze user behaviors in a non-intrusive and device-free manner, benefiting applications as diverse as smart homes and healthcare. However, most previous works focus on single-user sensing, which has limited practicability in scenarios involving multiple users. Although recent studies have begun to investigate WiFi-based multi-user sensing, there remains a lack of benchmark datasets to facilitate reproducible and comparable research. To bridge this gap, we present WiMANS, to our knowledge, the first dataset for multi-user sensing based on WiFi. WiMANS contains over 9.4 hours of dual-band WiFi Channel State Information (CSI), as well as synchronized videos, monitoring the simultaneous activities of multiple users. We exploit WiMANS to benchmark the performance of state-of-the-art WiFi-based human sensing models and video-based models, posing new challenges and opportunities for future work. We believe WiMANS can push the boundaries of current studies and catalyze the research on WiFi-based multi-user sensing.

We present EgoExo-Fitness, a new full-body action understanding dataset, featuring fitness sequence videos recorded from synchronized egocentric and fixed exocentric (third-person) cameras. Compared with existing full-body action understanding datasets, EgoExo-Fitness not only contains videos from first-person perspectives, but also provides rich annotations. Specifically, two-level temporal boundaries are provided to localize single action videos along with sub-steps of each action. More importantly, EgoExo-Fitness introduces innovative annotations for interpretable action judgement--including technical keypoint verification, natural language comments on action execution, and action quality scores. Combining all of these, EgoExo-Fitness provides new resources to study egocentric and exocentric full-body action understanding across dimensions of what'',when'', and how well''. To facilitate research on egocentric and exocentric full-body action understanding, we construct benchmarks on a suite of tasks (i.e., action recognition, action localization, cross-view sequence verification, cross-view skill determination, and a newly proposed task of guidance-based execution verification), together with detailed analysis. Data and code are available at https://github.com/iSEE-Laboratory/EgoExo-Fitness/tree/main.

We propose a simple yet effective approach for few-shot action recognition, emphasizing the disentanglement of motion and appearance representations. By harnessing recent progress in tracking, specifically point trajectories, and self-supervised representation learning, we build trajectory-aligned tokens (TATs) that capture motion and appearance information. This approach significantly reduces the data requirements while retaining essential information. To process these representations, we use a Masked Space-time Transformer that effectively learns to aggregate information to facilitate few-shot action recognition. We demonstrate state-of-the-art results on few-shot action recognition across multiple datasets.

Online Temporal Action Localization (On-TAL) is a critical task that aims to instantaneously identify action instances in untrimmed streaming videos as soon as an action concludes—a major leap from frame-based Online Action Detection (OAD). Yet, the challenge of detecting overlapping actions is often overlooked even though it is a common scenario in streaming videos. Current methods that can address concurrent actions depend heavily on class information, limiting their flexibility. This paper introduces ActionSwitch, the first class-agnostic On-TAL framework capable of detecting overlapping actions. By obviating the reliance on class information, ActionSwitch provides wider applicability to various situations, including overlapping actions of the same class or scenarios where class information is unavailable. This approach is complemented by the proposed "conservativeness loss", which directly embeds a conservative decision-making principle into the loss function for On-TAL. Our ActionSwitch achieves state-of-the-art performance in complex datasets, including Epic-Kitchens 100 targeting the challenging egocentric view and FineAction consisting of fine-grained atomic actions. Code will be made available.

Learning to infer labels in an open world, i.e., in an environment where the target "labels" are unknown, is an important characteristic for achieving autonomy. Foundation models, pre-trained on enormous amounts of data, have shown remarkable generalization skills through prompting, particularly in zero-shot inference. However, their performance is restricted to the correctness of the target label's search space, i.e., candidate labels provided in the prompt. This target search space can be unknown or exceptionally large in an open world, severely restricting their performance. To tackle this challenging problem, we propose a two-step, neuro-symbolic framework called ALGO - Action Learning with Grounded Object recognition that uses symbolic knowledge stored in large-scale knowledge bases to infer activities in egocentric videos with limited supervision. First, we propose a neuro-symbolic prompting approach that uses object-centric vision-language models as a noisy oracle to ground objects in the video through evidence-based reasoning. Second, driven by prior commonsense knowledge, we discover plausible activities through an energy-based symbolic pattern theory framework and learn to ground knowledge-based action (verb) concepts in the video. Extensive experiments on four publicly available datasets (EPIC-Kitchens, GTEA Gaze, GTEA Gaze Plus, and Charades-Ego) demonstrate its performance on open-world activity inference. We also show that …

A novel algorithm for video lane detection is proposed in this paper. First, we extract a feature map for a current frame and detect a latent mask for obstacles occluding lanes. Then, we enhance the feature map by developing an occlusion-aware memory-based refinement (OMR) module. It takes the obstacle mask and feature map from the current frame, previous output, and memory information as input, and processes them recursively in a video. Moreover, we apply a novel data augmentation scheme for training the OMR module effectively. Experimental results show that the proposed algorithm outperforms existing techniques on video lane datasets. The source codes will be made publicly available.

Modern video segmentation methods adopt object queries to perform inter-frame association and demonstrate satisfactory performance in tracking continuously appearing objects despite large-scale motion and transient occlusion. However, they all underperform on newly emerging and disappearing objects that are common in the real world because they attempt to model object emergence and disappearance through feature transitions between background and foreground queries that have significant feature gaps. We introduce Dynamic Anchor Queries (DAQ) to shorten the transition gap between the anchor and target queries by dynamically generating anchor queries based on the features of potential candidates. Furthermore, we introduce a query-level object Emergence and Disappearance Simulation (EDS) strategy, which unleashes the potential of DAQ without any additional cost. Finally, we combine our proposed DAQ and EDS with DVIS~\cite{zhang2023dvis} to obtain DVIS-DAQ. Extensive experiments demonstrate that DVIS-DAQ achieves a new SOTA performance on five mainstream video segmentation benchmarks. Code and models will be available for further study.

In recent years, online Video Instance Segmentation (VIS) methods have shown remarkable advancement with their powerful query-based detectors. Utilizing the output queries of the detector at the frame-level, these methods achieve high accuracy on challenging benchmarks. However, our observations demonstrate that these methods heavily rely on location information, which often causes incorrect associations between objects. This paper presents that a key axis of object matching in trackers is appearance information, which becomes greatly instructive under conditions where positional cues are insufficient for distinguishing their identities. Therefore, we suggest a simple yet powerful extension to object decoders that explicitly extract embeddings from backbone features and drive queries to capture the appearances of objects, which greatly enhances instance association accuracy. Furthermore, recognizing the limitations of existing benchmarks in fully evaluating appearance awareness, we have constructed a synthetic dataset to rigorously validate our method. By effectively resolving the over-reliance on location information, we achieve state-of-the-art results on YouTube-VIS 2019/2021 and Occluded VIS (OVIS). Code is available at https://github.com/KimHanjung/VISAGE.

Humans can foresee the future based on present observations, a skill we term as foresight minds. However, this capability remains under-explored within existing MLLMs, hindering their capacity to understand intentions behind subjects. To address this, we integrate the future modeling into MLLMs. By utilizing the trajectory, a highly structured representation, as a learning objective, we aim to equip the model to understand spatiotemporal dynamics. Inspired by the learning paradigm of LLMs, we first propose Foresight Pre-Training (FPT) that jointly learns various tasks centered on trajectories, enabling MLLMs to predict entire trajectories from a given initial observation. Then, we propose Foresight Instruction-Tuning (FIT) that requires MLLMs to reason about potential future events based on predicted trajectories. Aided by FPT and FIT, we build an unified MLLM named Merlin that supports complex future reasoning. Experiments show Merlin’s foresight minds with impressive performance on both future reasoning and visual comprehension tasks.

Video class-incremental learning (VCIL) aims to learn discriminative and generalized feature representations for video frames to mitigate catastrophic forgetting. Conventional VCIL approaches often retain a subset of frames or features from prior tasks as exemplars for subsequent incremental learning stages. However, these strategies overlook the connection between base and novel classes, sometimes even leading to privacy leakage. To overcome this challenge, we introduce a Spatial-Temporal Subspace Projection (STSP) scheme for VCIL. Specifically, we propose a discriminative Temporal-based Subspace Classifier (TSC) that represents each class with an orthogonal subspace basis and adopts subspace projection loss for classification. Differing from typical classification methods that rely on fully connected layers, our TSC is designed to discern the spatial-temporal dynamics in video content, thereby enhancing the representation of each video sample. Additionally, we implement inter- and intra-class orthogonal constraints into TSC, ensuring that each class occupies a unique orthogonal subspace, defined by its basis. To prevent catastrophic forgetting, we further employ a Spatial-based Gradient Projection (SGP) strategy. SGP adjusts the gradients of the network parameters to align with the approximate null space of the spatial feature set from previous tasks. Extensive experiments conducted on three benchmarks, namely HMDB51, UCF101, and Something-Something V2, demonstrate that …

Temporal Action Detection (TAD) focuses on detecting pre-defined actions, while Moment Retrieval (MR) aims to identify the events described by open-ended natural language within untrimmed videos. Despite that they focus on different events, we observe they have a significant connection. For instance, most descriptions in MR involve multiple actions from TAD. In this paper, we aim to investigate the potential synergy between TAD and MR. Firstly, we propose a unified architecture, termed Unified Moment Detection (UniMD), for both TAD and MR. It transforms the inputs of the two tasks, namely actions for TAD or events for MR, into a common embedding space, and utilizes two novel query-dependent decoders to generate a uniform output of classification score and temporal segments. Secondly, we explore the efficacy of two task fusion learning approaches, pre-training and co-training, in order to enhance the mutual benefits between TAD and MR. Extensive experiments demonstrate that the proposed task fusion learning scheme enables the two tasks to help each other and outperform the separately trained counterparts. Impressively, UniMD achieves state-of-the-art results on three paired datasets Ego4D, Charades-STA, and ActivityNet.

Video summarization plays a vital role in improving video browsing efficiency and has various applications in action recognition and information retrieval. In order to generate summaries that can provide key information, existing works have been proposed to simultaneously explore the contribution of both long-range and short-range temporal cues. However, they rarely consider the potential correspondence between temporal cues at different granularity within video sequences, making it insufficient to ensure detailed video understanding. In order to solve this issue, we propose a novel video summarization framework, namely Bgm4Video, based on the graph-matching mechanism, which models the potential contextualized relationship across multi-granularity temporal cues. The proposed framework is composed of two dominant components including (i) a temporal encoder (TE) that explores both coarse-grained and fine-grained contextual information within videos and (ii) a bidirectional graph transmission (BGT) module that models the interrelationship across multi-granularity temporal cues. By grasping the contextual correspondence, our method allows for further refining temporal representations to precisely pinpoint valuable segments. We demonstrate the advantage of our components through an extensive ablation study. We also evaluate our full approach on the video summarization task and demonstrate improvements over state-of-the-art on the popular benchmarks.

This paper explores the spatio-temporal video grounding (STVG) task, which aims at localizing a particular object corresponding to a given textual description in an untrimmed video. Existing approaches mainly resort to object-level manual annotations as the supervision for addressing this challenging task. Such a paradigm heavily constrains the scalability of processing large-scale unlabeled data. To this end, we present a novel framework that is capable of grounding the target object relying only on the video-sentence correspondence. Specifically, our model re-formulates the original STVG task as two cross-modal alignment sub-problems: region-phrase and frame-sentence. Since the absence of ground-truth alignments during the training stage, we treat them as latent variables and learn to model the joint conditional distribution by reconstructing the interactions of entities in the video. The entire framework can be effectively optimized by the variational Expectation-Maximization (EM) algorithm, which alternates between two updating steps for progressively maximizing the likelihood of query sentence, thereby approximating the real cross-modal assignment. Extensive experiments on two video benchmarks (VidSTG and HC-STVG) further show the effectiveness of the proposed method.

Egocentric videos provide a unique perspective into individuals' daily experiences, yet their unstructured nature presents challenges for perception. In this paper, we introduce AMEGO, a novel approach aimed at enhancing the comprehension of very-long egocentric videos. Inspired by the human's ability to memorise information from a single watching, our method focuses on constructing self-contained representations from the egocentric video, capturing key locations and object interactions. This representation is semantic-free and facilitates multiple queries without the need to reprocess the entire visual content. Additionally, to evaluate our understanding of very-long egocentric videos, we introduce the new Active Memories Benchmark (AMB), composed of more than 20K of highly challenging visual queries from EPIC-KITCHENS. These queries cover different levels of video reasoning (sequencing, concurrency and temporal grounding) to assess detailed video understanding capabilities. We showcase improved performance of AMEGO on AMB, surpassing other video QA baselines by a substantial margin.

This paper addresses the challenging task of weakly-supervised video temporal grounding. Existing approaches are generally based on the moment candidate selection pipeline that utilizes contrastive learning and reconstruction paradigm for scoring the pre-defined moments. Although they have achieved significant progress, we argue that their current frameworks have overlooked two indispensable issues: (1) Coarse-grained cross-modal learning: previous methods solely capture the global video-level alignment with the query, failing to model the detailed consistency between video frames and query words for accurately grounding the moment boundaries. (2) Complex moment candidates: the performance of these methods severely relies on the quality of moment candidates, which are also time-consuming and complicated for selection. To this end, in this paper, we make the first attempt to tackle this task from a novel game perspective, which effectively learns the uncertain relationship between each frame-word pair with diverse granularity and flexible combination for fine-grained cross-modal interaction. Specifically, we creatively model each video frame and query word as game players with multivariate cooperative game theory to learn their contribution to the cross-modal similarity score. By quantifying the trend of frame-word cooperation within a coalition via the game-theoretic interaction, we are able to value all uncertain but possible correspondence …

Video reasoning typically operates within the Video Question-Answering (VQA) paradigm, which demands that the models understand and reason about video content from temporal and causal perspectives. Traditional supervised VQA methods gain this capability through meticulously annotated QA datasets, while advanced visual-language models exhibit remarkable performance due to large-scale visual-text pretraining data. Nevertheless, due to potential language bias and spurious visual-text correlations in cross-modal learning, concerns about the reliability of their answers persist in real-world applications. In this paper, we focus on the grounded VQA task, which necessitates models to provide answers along with explicit visual evidence, i.e., certain video segments. As temporal annotation is not available during training, we propose a novel bi-directional reasoning framework to perform grounded VQA in a weakly-supervised setting. Specifically, our framework consists of two parallel but dual reasoning paths. They conduct temporal grounding and answering based on the video content, approaching it from two dual directions that are symmetrical in terms of temporal order or causal relationships. By constructing a cycle-consistency relationship between these two branches, the model is prompted to provide self-guidance supervision for both temporal grounding and answering. Experiments conducted on the Next-GQA and Env-QA datasets demonstrate that our framework achieves superior performance …

Understanding and modeling the popularity of User Generated Content (UGC) short videos on social media platforms presents a critical challenge with broad implications for content creators and recommendation systems. This study delves deep into the intricacies of predicting engagement for newly published videos with limited user interactions. Surprisingly, our findings reveal that Mean Opinion Scores from previous video quality assessement datasets do not strongly correlate with video engagement levels. To address this, we introduce a substantial dataset comprising 90,000 real-world UGC short videos from Snapchat. Rather than relying on view count, average watch time, or rate of likes, we propose two metrics: normalized average watch percentage (NAWP) and engagement continuation rate (ECR) to describe the engagement levels of short videos. Comprehensive multi-modal features, including visual content, background music, and text data, are investigated to enhance engagement prediction. With the proposed dataset and two key metrics, our method demonstrates its ability to predict engagements of short videos purely from video content.

There has been tremendous progress in multimodal Large Language Models (LLMs). Recent works have extended these models to video input with promising instruction following capabilities. However, an important missing piece is temporal localization. These models cannot accurately answer the "When?" questions. We identify three key aspects that limit their temporal localization capabilities: (i) time representation, (ii) architecture, and (iii) data. We address these shortcomings by proposing Language Instructed Temporal-Localization Assistant (LITA) with the following features: (1) We introduce time tokens that encode timestamps relative to the video length to better represent time in videos. (2) We introduce SlowFast tokens in the architecture to capture temporal information at fine temporal resolution. (3) We emphasize temporal localization data for LITA. In addition to leveraging existing video datasets with timestamps, we propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for learning and evaluating this task. Reasoning temporal localization requires both the reasoning and temporal localization of Video LLMs. LITA demonstrates strong performance on this challenging task, nearly doubling the temporal mean intersection-over-union (mIoU) of baselines. In addition, we show that our emphasis on temporal localization also substantially improves video-based text generation compared to existing Video LLMs, including a …

Weakly supervised audio-visual video parsing (AVVP) methods aim to detect audible-only, visible-only, and audible-visible events using only video-level labels. Existing approaches tackle this by leveraging both unimodal and cross-modal contexts. However, we argue that while cross-modal learning is beneficial for detecting audible-visible events, in the weakly supervised scenario, it negatively impacts unaligned audible or visible events by introducing irrelevant modality information. In this paper, we propose CoLeaF, a novel learning framework that optimizes the integration of cross-modal context in the embedding space such that the network explicitly learns to combine cross-modal information for audible-visible events while filtering them out for unaligned events. Additionally, as videos often involve complex class relationships, modelling them improves performance. However, this introduces extra computational costs into the network. Our framework is designed to leverage cross-class relationships during training without incurring additional computations at inference. Furthermore, we propose new metrics to better evaluate a method's capabilities in performing AVVP. Our extensive experiments demonstrate that CoLeaF significantly improves the state-of-the-art results by an average of 1.9% and 2.4% F-score on the LLP and UnAV-100 datasets, respectively. Our code will be released upon paper publication.

Traditional audio-visual methods rely on independent audio and visual backbones, which is costly and not scalable. In this work, we investigate using an audio-visual siamese network (AVSiam) for efficient and scalable audio-visual pretraining. Our framework uses a single shared vision transformer backbone to process audio and visual inputs, improving its parameter efficiency, reducing the GPU memory footprint, and allowing us to scale our method to larger datasets and model sizes. We pretrain our model using a contrastive audio-visual matching objective with a multi-ratio random masking scheme, which further speeds up the audio-visual pretraining process and enables our model to process larger audio-visual instance batches, helpful for contrastive learning. Unlike prior audio-visual methods, our method can robustly handle audio-only, visual-only, and audio-visual inputs with a single shared ViT backbone. Furthermore, despite using the shared backbone for both modalities, AVSiam achieves competitive or even better results than prior methods on AudioSet and VGGSound for audio-visual classification and audio-visual retrieval. Our code and models will be made publicly available.

Sign language is one of the most effective communication tools for people with hearing difficulties. Most existing works focus on improving the performance of sign language tasks on RGB videos, which may suffer from degraded recording conditions, such as fast movement of hands with motion blur and textured signer’s appearance. The bio-inspired event camera, which asynchronously captures brightness change with high speed, could naturally perceive dynamic hand movements, providing rich manual clues for sign language tasks. In this work, we aim at exploring the potential of event camera in continuous sign language recognition (CSLR) and sign language translation (SLT). To promote the research, we first collect an event-based benchmark EvSign for those tasks with both gloss and spoken language annotations. EvSign dataset offers a substantial amount of high-quality event streams and an extensive vocabulary of glosses and words, thereby facilitating the development of sign language tasks. In addition, we propose an efficient transformer-based framework for event-based SLR and SLT tasks, which fully leverages the advantages of streaming events. The sparse backbone is employed to extract visual features from sparse events. Then, the temporal coherence is effectively utilized through the proposed local token fusion and gloss-aware temporal aggregation modules. Extensive experimental …

In this paper, we address the challenge of fine-grained video event understanding in traffic scenarios, vital for autonomous driving and safety. Traditional datasets focus on driver or vehicle behavior, often neglecting pedestrian perspectives. To fill this gap, we introduce the WTS dataset, highlighting detailed behaviors of both vehicles and pedestrians across over 1.2k video events in over hundreds traffic scenarios. WTS integrates diverse perspectives from vehicle ego and fixed overhead cameras in a vehicle-infrastructure cooperative environment, enriched with comprehensive textual descriptions and unique 3D Gaze data for a synchronized 2D/3D view, focusing on pedestrian analysis. We also provide annotations for 5k publicly sourced pedestrian-related traffic videos. Additionally, we introduce LLMScorer, an LLM-based evaluation metric to align inference captions with ground truth. Using WTS, we establish a benchmark for dense video-to-text tasks, exploring state-of-the-art Vision-Language Models with an instance-aware VideoLLM method as a baseline. WTS aims to advance fine-grained video event understanding, enhancing traffic safety and autonomous driving development. Dataset page: https://woven-visionai.github.io/wts-dataset-homepage/.

Unsupervised visible-infrared person re-identification (USL-VI-ReID) is a promising yet highly challenging retrieval task. The key challenges in USL-VI-ReID are to accurately generate pseudo-labels and establish pseudo-label correspondences across modalities without relying on any prior annotations. Recently, clustered pseudo-label methods have gained more attention in USL-VI-ReID. However, most existing methods don't fully exploit the intra-class nuances, as they simply utilize a single memory that represents an identity to establish cross-modality correspondences, resulting in noisy cross-modality correspondences. To address the problem, we propose a Multi-Memory Matching (MMM) framework for USL-VI-ReID. We first design a simple yet effective Cross-Modality Clustering (CMC) module to generate the pseudo-labels through clustering together both two modality samples. To associate cross-modality clustered pseudo-labels, we design a Multi-Memory Learning and Matching (MMLM) module, ensuring that optimization explicitly focuses on the nuances of individual perspectives and establishes reliable cross-modality correspondences. Finally, we design a Soft Cluster-level Alignment (SCA) loss to narrow the modality gap while mitigating the effect of noisy pseudo-labels through a soft many-to-many alignment strategy. Extensive experiments on the public SYSU-MM01 and RegDB datasets demonstrate the reliability of the established cross-modality correspondences and the effectiveness of MMM. The source codes will be released.

To overcome the inherent domain gap between remote sensing (RS) images and natural images, some self-supervised representation learning methods have made promising progress. However, they have overlooked the diverse angles present in RS objects. This paper proposes the Masked Angle-Aware Autoencoder (MA3E) to perceive and learn angles during pre-training. We design a scaling center crop operation to create the rotated crop with random orientation on each original image, introducing the explicit angle variation. MA3E inputs this composite image while reconstruct the original image, aiming to effectively learn rotation-invariant representations by restoring the angle variation introduced on the rotated crop. To avoid biases caused by directly reconstructing the rotated crop, we propose an Optimal Transport (OT) loss that automatically assigns similar original image patches to each rotated crop patch for reconstruction. MA3E demonstrates more competitive performance than existing pre-training methods on seven different RS image datasets in three downstream tasks.

Accurately recognizing a revisited place is crucial for embodied agents to localize and navigate. This requires visual representations to be distinct, despite strong variations in camera viewpoint and scene appearance. Existing visual place recognition pipelines encode the whole' image and search for matches. This poses a fundamental challenge in matching two images of the same place captured from different camera viewpoints: "the similarity of what overlaps can be dominated by the dissimilarity of what does not overlap". We address this by encoding and searching forimage segments' instead of the whole images. We propose to use open-set image segmentation to decompose an image into meaningful' entities (i.e., things and stuff). This enables us to create a novel image representation as a collection of multiple overlapping subgraphs connecting a segment with its neighboring segments, dubbed SuperSegment. Furthermore, to efficiently encode these SuperSegments into compact vector representations, we propose a novel factorized representation of feature aggregation. We show that retrieving these partial representations leads to significantly higher recognition recall than the typical whole image based retrieval. Our segments-based approach, dubbed SegVLAD, sets a new state-of-the-art in place recognition on a diverse selection of benchmark datasets, while being applicable toboth' generic and …

Embodied visual tracking is to follow a target object in dynamic 3D environments using an agent’s egocentric vision. This is a vital and challenging skill for embodied vision agents. However, existing methods suffer from inefficient training and poor generalization. In this paper, we propose a novel framework that combines visual foundation models (VFM) and offline reinforcement learning (offline RL) to empower embodied visual tracking. We use a pre-trained VFM, such as Tracking Anything”, to extract semantic segmentation masks with text prompts. We then train a recurrent policy network with offline RL, e.g., Conservative Q-Learning, to learn from the collected demonstrations without online agent-environment interactions. To improve the training efficiency, robustness, and generalization of the policy network, we also introduce a mask re-targeting mechanism and a multi-level data collection strategy. In this way, we can train the tracker within an hour on a consumer-level GPU, e.g., Nvidia RTX 3090. Such efficiency is unprecedented for RL-based visual tracking methods. We evaluate our tracker on several high-fidelity environments with challenging situations, such as distraction and occlusion. The results show that our agent outperforms state-of-the-art methods in terms of sample efficiency, robustness to distractors, and generalization to unseen scenarios and targets.

One of the essential missions in the AI research community is to build an autonomous embodied agent that can achieve high-level performance across a wide spectrum of tasks. However, acquiring or manually designing rewards for all open-ended tasks is unrealistic. In this paper, we propose a novel cross-modal contrastive learning framework architecture, CLIP4MC, aiming to learn a reinforcement learning (RL) friendly vision-language model (VLM) that serves as an intrinsic reward function for open-ended tasks. Simply utilizing the similarity between the video snippet and the language prompt is not RL-friendly since standard VLMs may only capture the similarity at a coarse level. To achieve RL-friendliness, we incorporate the task completion degree into the VLM training objective, as this information can assist agents in distinguishing the importance between different states. Moreover, we provide neat YouTube datasets based on the large-scale YouTube database provided by MineDojo. Specifically, two rounds of filtering operations guarantee that the dataset covers enough essential information and that the video-text pair is highly correlated. Empirically, we demonstrate that the proposed method achieves better performance on RL tasks compared with baselines.

Building a general-purpose intelligent home-assistant agent skilled in diverse tasks by human commands is a long-term blueprint of embodied AI research, which poses requirements on task planning, environment modeling, and object interaction. In this work, we study primitive mobile manipulations for embodied agents, i.e. how to navigate and interact based on an instructed verb-noun pair. We propose DISCO, which features non-trivial advancements in contextualized scene modeling and efficient controls. In particular, DISCO incorporates differentiable scene representations of rich semantics in object and affordance, which is dynamically explored on the fly and facilitates navigation planning. Besides, we propose dual-level coarse-to-fine action controls leveraging both global and local cues to accomplish mobile manipulation tasks efficiently. DISCO easily integrates into embodied tasks such as embodied instruction following. To validate our approach, we take the ALFRED benchmark, of large-scale long-horizon vision-language navigation and interaction tasks, as a test bed. In extensive experiments, we make comprehensive evaluations and demonstrate that DISCO outperforms the art by a sizable +8.6\% success rate margin in unseen scenes, even without step-by-step instructions. Our code and model will be made publicly available.

Large language models (LLMs) have achieved impresx0002sive progress on several open-world tasks. Recently, usx0002ing LLMs to build embodied agents has been a hotspot. In this paper, we propose STEVE, a comprehensive and visionx0002ary embodied agent in the Minecraft virtual environment. STEVE consists of three key components: vision perception, language instruction, and code action. Vision perception involves the interpretation of visual information in the envix0002ronment, which is then integrated into the LLMs component with agent state and task instruction. Language instrucx0002tion is responsible for iterative reasoning and decomposx0002ing complex tasks into manageable guidelines. Code action generates executable skill actions based on retrieval in skill database, enabling the agent to interact effectively within the Minecraft environment. We also collect STEVE-21K dataset, which includes 600+ vision-environment pairs, 20K knowledge question-answering pairs, and 200+ skillx0002code pairs. We conduct continuous block search, knowledge question and answering, and tech tree mastery to evaluate the performance. Extensive experiments show that STEVE achieves at most 1.5× faster unlocking key tech trees and 2.5× quicker in block search tasks compared to previous state-of-the-art methods.

Aligning multiple modalities in a latent space, such as images and texts, has shown to produce powerful semantic visual representations, fueling tasks like image captioning, text-to-image generation, or image grounding. In the context of human-centric vision, albeit these CLIP-like representations encode most standard human poses relatively well (such as standing or sitting), they lack sufficient acuteness to discern detailed or uncommon ones. Actually, while 3D human poses have been often associated with images (e.g. to perform pose estimation or pose-conditioned image generation), or more recently with text (e.g. for text-to-pose generation), they have seldom been paired with both. In this work, we combine 3D poses, person's pictures and textual pose descriptions to produce an enhanced 3D-, visual- and semantic-aware human pose representation. We introduce a new transformer-based model, trained in a retrieval fashion, which can take as input any combination of the aforementioned modalities. When composing modalities, it outperforms a standard multi-modal alignment retrieval model, making it possible to sort out partial information (e.g. image with the lower body occluded). We showcase the potential of such an embroidered pose representation on the task of fine-grained instruction generation, which consists in generating a text that describes how to move from one …

Recent advances in visual reasoning (VR), particularly with the aid of Large Vision-Language Models (VLMs), show promise but require access to large-scale datasets and face challenges such as high computational costs and limited generalization capabilities. Compositional visual reasoning approaches have emerged as effective strategies; however, they heavily rely on the commonsense knowledge encoded in Large Language Models (LLMs) to perform planning, reasoning, or both, without considering the effect of their decisions on the visual reasoning process, which can lead to errors or failed procedures. To address these challenges, we introduce HYDRA, a multi-stage dynamic compositional visual reasoning framework designed for reliable and incrementally progressive general reasoning. HYDRA integrates three essential modules: a planner, a Reinforcement Learning (RL) agent serving as a cognitive controller, and a reasoner. The planner and reasoner modules utilize an LLM to generate instruction samples and executable code from the selected instruction, respectively, while the RL agent dynamically interacts with these modules, making high-level decisions on selection of the best instruction sample given information from the historical state stored through a feedback loop. This adaptable design enables HYDRA to adjust its actions based on previous feedback received during the reasoning process, leading to more reliable reasoning outputs …

Visual reasoning, as a prominent research area, plays a crucial role in AI by facilitating concept formation and interaction with the world. However, current works are usually carried out separately on small datasets thus lacking generalization ability. Through rigorous evaluation on diverse benchmarks, we demonstrate the shortcomings of existing methods in achieving cross-domain reasoning and their tendency to data bias fitting. In this paper, we revisit visual reasoning with a two-stage perspective: (1) symbolization and (2) logical reasoning given symbols or their representations. We find that the reasoning stage is better at generalization than symbolization. Thus, it is more efficient to implement symbolization via separated encoders for different data domains while using a shared reasoner. Given our findings, we establish design principles for visual reasoning frameworks following the separated symbolization and shared reasoning. Our two-stage framework achieves impressive generalization ability on various visual reasoning tasks, including puzzles, physical prediction, and visual question answering (VQA), encompassing both 2D and 3D modalities. We believe our insights will pave the way for generalizable visual reasoning. Our code will be publicly available.

The existing works on object-level language grounding with 3D objects mostly focus on improving performance by utilizing the off-the-shelf pre-trained models to capture features, such as viewpoint selection or geometric priors. However, they have failed to consider exploring the cross-modal representation of language-vision alignment in the cross-domain field. To answer this problem, we propose a novel method called Domain Adaptation for Language Grounding (DA4LG) with 3D objects. Specifically, the proposed DA4LG consists of a visual adapter module with multi-task learning to realize vision-language alignment by comprehensive multimodal feature representation. Experimental results demonstrate that DA4LG competitively performs across visual and non-visual language descriptions, independent of the completeness of observation. DA4LG achieves state-of-the-art performance in the single-view setting and multi-view setting with the accuracy of 83.8 % and 86.8 % respectively in the language grounding benchmark SNARE. The simulation experiments show the well-practical and generalized performance of DA4LG compared to the existing methods. Our project is available anonymously at https://sites.google.com/view/da4lg.

The remarkable progress of Multi-modal Large Language Models (MLLMs) has garnered unparalleled attention, due to their superior performance in visual contexts. However, their capabilities in visual math problem-solving remain insufficiently evaluated and understood. We investigate current benchmarks to incorporate excessive visual content within textual questions, which potentially assist MLLMs in deducing answers without truly interpreting the input diagrams. To this end, we introduce MathVerse, an all-around visual math benchmark designed for an equitable and in-depth evaluation of MLLMs. We meticulously collect 2,612 high-quality, multi-subject math problems with diagrams from publicly available sources. Each problem is then transformed by human annotators into six distinct versions, each offering varying degrees of information content in multi-modality, contributing to 15K test samples in total. This approach allows MathVerse to comprehensively assess whether and how much MLLMs can truly understand the visual diagrams for mathematical reasoning. In addition, we propose a Chain-of-Thought (CoT) evaluation strategy for a fine-grained assessment of the output answers. Rather than naively judging true or false, we employ GPT-4(V) to adaptively extract crucial reasoning steps, and then assess each step with error analysis to derive a total score, which can reveal the inner CoT reasoning quality by MLLMs. With MathVerse, we …

With the breakthrough of multi-modal large language models, answering complex visual questions that demand advanced reasoning abilities and world knowledge has become a much more important testbed for developing AI models than ever. However, equipping AI models with robust cross-modality reasoning ability remains challenging since the cognition scheme of humans has not been understood systematically. In this paper, we believe that if we can collect visual clues of each instance in the given image, we will recognize the image more accurately, understand the question better, recall relevant knowledge more easily, and finally reason out the answer. We discover these important and rich visual clues by mining question-answer pairs in images and sending them into multi-modal large language models as prompts. We call the proposed method Q&A Prompts. Specifically, we first use the image-answer pairs and the corresponding questions in the training set as inputs and outputs to train a visual question generation model. Then, we use an image tagging model to identify various instances and send packaged image-tag pairs into the visual question generation model to generate relevant questions with the extracted image tags as answers. Finally, we encode these generated question-answer pairs as prompts with a visual-aware prompting module …

With the recent significant advancements in large multimodal models (LMMs), the importance of their grounding capability in visual chat is increasingly recognized. Despite recent efforts to enable LMMs to support grounding, their capabilities for grounding and chat are usually separate, and their chat performance drops dramatically when asked to ground. The problem is the lack of a dataset for grounded visual chat (GVC). Existing grounding datasets only contain short captions. To address this issue, we have created GVC data that allows for the combination of grounding and chat capabilities. To better evaluate the GVC capabilities, we have introduced a benchmark called Grounding Bench. Additionally, we have proposed a model design that can support GVC and various types of visual prompts by connecting segmentation models with language models. Experimental results demonstrate that our model outperforms other LMMs on Grounding Bench. Furthermore, our model achieves competitive performance on classic grounding benchmarks like RefCOCO/+/g and Flickr30K Entities.

This work focuses on the potential of vision large language models (VLLMs) in visual reasoning. Different from prior studies, we shift our focus from evaluating standard performance to introducing a comprehensive safety evaluation suite Unicorn, covering out-of-distribution (OOD) generalization and adversarial robustness. For the OOD evaluation, we present two novel visual question-answering (VQA) datasets, each with one variant, designed to test model performance under challenging conditions. In exploring adversarial robustness, we propose a straightforward attack strategy for misleading VLLMs to produce visual-unrelated responses. Moreover, we assess the efficacy of two jailbreaking strategies, targeting either the vision or language input of VLLMs. Our evaluation of 22 diverse models, ranging from open-source VLLMs to GPT-4V and Gemini Pro, yields interesting observations: 1) Current VLLMs struggle with OOD texts but not images, unless the visual information is limited; and 2) These VLLMs can be easily misled by deceiving vision encoders only, and their vision-language training often compromise safety protocols. We release this safety evaluation suite at https://github.com/UCSC-VLAA/vllm-safety-benchmark.

Warning: This paper contains examples of harmful language and images, and reader discretion is recommended. The security concerns surrounding Large Language Models (LLMs) have been extensively explored, yet the safety of Multimodal Large Language Models (MLLMs) remains understudied. In this paper, we observe that Multimodal Large Language Models (MLLMs) can be easily compromised by query-relevant images, as if the text query itself were malicious. To address this, we introduce MM-SafetyBench, a comprehensive framework designed for conducting safety-critical evaluations of MLLMs against such image-based manipulations. We have compiled a dataset comprising 13 scenarios, resulting in a total of 5,040 text-image pairs. Our analysis across 12 state-of-the-art models reveals that MLLMs are susceptible to breaches instigated by our approach, even when the equipped LLMs have been safety-aligned. In response, we propose a straightforward yet effective prompting strategy to enhance the resilience of MLLMs against these types of attacks. Our work underscores the need for a concerted effort to strengthen and enhance the safety measures of open-source MLLMs against potential malicious exploits.

Vision-language pre-training (VLP) models exhibit remarkable capabilities in comprehending both images and text, yet they remain susceptible to multimodal adversarial examples (AEs). Strengthening adversarial attacks and uncovering vulnerabilities, especially common issues in VLP models (e.g., high transferable AEs), can stimulate further research on constructing reliable and practical VLP models. A recent work (i.e., Set-level guidance attack) indicates that augmenting image-text pairs to increase AE diversity along the optimization path enhances the transferability of adversarial examples significantly. However, this approach predominantly emphasizes diversity around the online adversarial examples (i.e., AEs in the optimization period), leading to the risk of overfitting the victim model and affecting the transferability. In this study, we posit that the diversity of adversarial examples towards the clean input and online AEs are both pivotal for enhancing transferability across VLP models. Consequently, we propose using diversification along the intersection region of adversarial trajectory to expand the diversity of AEs. To fully leverage the interaction between modalities, we introduce text-guided adversarial example selection during optimization. Furthermore, to further mitigate the potential overfitting, we direct the adversarial text deviating from the last intersection region along the optimization path, rather than adversarial images as in existing methods. Extensive experiments affirm the …

Modality alignment has been of paramount importance in recent developments of multimodal learning, which has inspired many innovations in multimodal networks and pre-training tasks. Single-stream networks can effectively leverage self-attention mechanisms to facilitate modality interactions but suffer from high computational complexity and limited applicability to downstream retrieval tasks. In contrast, dual-stream networks address these issues but ignore the significance of modality alignment. In this paper, we develop a multimodal learning method that integrates the advantages of modality alignment from single-stream networks into the dual-stream network by introducing object-oriented anchors to bridge alignment between image and text modalities. Object-oriented anchors are generated effectively and circumvent the need for object detection boxes as previous region-based approaches, while also preserving explicit semantics for modality interactions. Additionally, we design fine-grained token-level asymmetry alignment between modalities and multiview mining to promote modality alignment. To the best of our knowledge, we are the first to apply object-oriented tokens in multimodal pre-training, yielding significant benefits. Extensive experimental results validate the effectiveness of our method, demonstrating that the proposed method outperforms most previous arts in various downstream tasks, particularly when considering comparable data and model scales.

Most advanced visual grounding methods rely on Transformers for visual-linguistic feature fusion. However, these Transformer-based approaches encounter a significant drawback: the computational costs escalate quadratically due to the self-attention mechanism in the Transformer Encoder, particularly when dealing with high-resolution images or long context sentences. This quadratic increase in computational burden restricts the applicability of visual grounding to more intricate scenes, such as conversation-based reasoning segmentation, which involves lengthy language expressions. In this paper, we propose an efficient and effective multi-task visual grounding (EEVG) framework based on Transformer Decoder to address this issue, which reduces the cost in both language and visual aspects. In the language aspect, we employ the Transformer Decoder to fuse visual and linguistic features, where linguistic features are input as memory and visual features as queries. This allows fusion to scale linearly with language expression length. In the visual aspect, we introduce a parameter-free approach to reduce computation by eliminating background visual tokens based on attention scores. We then design a light mask head to directly predict segmentation masks from the remaining sparse feature maps. Extensive results and ablation studies on benchmarks demonstrate the efficiency and effectiveness of our approach.

Hallucinations in vision-language models pose a significant challenge to their reliability, particularly in the generation of long captions. Current methods fall short of accurately identifying and mitigating these hallucinations. To address this issue, we introduce ESREAL, a novel unsupervised learning framework designed to suppress the generation of hallucinations through accurate localization and penalization of hallucinated tokens. Initially, ESREAL creates a reconstructed image based on the generated caption and aligns its corresponding regions with those of the original image. This semantic reconstruction aids in identifying both the presence and type of token-level hallucinations within the generated caption. Subsequently, ESREAL computes token-level hallucination scores by assessing the semantic similarity of aligned regions based on the type of hallucination. Finally, ESREAL employs a proximal policy optimization algorithm, where it selectively penalizes hallucinated tokens according to their token-level hallucination scores. Our framework notably reduces hallucinations in LLaVA, InstructBLIP, and mPLUG-Owl2 by 32.81%, 27.08%, and 7.46% on the CHAIR metric. This improvement is achieved solely through signals derived from the image itself, without the need for any image-text pairs.

Vision-language foundation models, represented by Contrastive Language-Image Pre-training (CLIP), have gained increasing attention for jointly understanding both vision and textual tasks. However, existing approaches primarily focus on training models to match global image representations with textual descriptions, thereby overlooking the critical alignment between local regions and corresponding text tokens. This paper extends CLIP with multi-granularity alignment. Notably, we deliberately construct a new dataset comprising pseudo annotations at various levels of granularities, encompassing image-level, region-level as well as pixel-level captions and tags. Accordingly, we develop a Unified Multi-Granularity learning framework, termed UMG-CLIP, which simultaneously empowers the model with versatile perception abilities across different levels of detail. With parameter efficient tuning, UMG-CLIP surpasses current widely used CLIP variants and achieves state-of-the-art performance on diverse image understanding benchmarks, including open-world recognition, retrieval, semantic segmentation, and panoptic segmentation tasks. We believe that UMG-CLIP represents a valuable advancement in vision-language foundation models.

When an image generation process is guided by both a text prompt and spatial cues, such as a set of bounding boxes, do these elements work in harmony, or does one dominate the other? Our analysis of a pretrained image diffusion model that integrates gated self-attention into the U-Net reveals that spatial grounding often outweighs textual grounding due to the sequential flow from gated self-attention to cross-attention. We demonstrate that such bias can be significantly mitigated without sacrificing accuracy in either grounding by simply rewiring the network architecture, changing from sequential to parallel for gated self-attention and cross-attention. This surprisingly simple yet effective solution does not require any fine-tuning of the network but significantly reduces the trade-off between the two groundings. Our experiments demonstrate significant improvements from the original GLIGEN to the rewired version in the trade-off between textual grounding and spatial grounding.

Reading text from images (either natural scenes or documents) has been a long-standing research topic for decades, due to the high technical challenge and wide application range. Previously, individual specialist models are developed to tackle the sub-tasks of text reading (e.g., scene text recognition, handwritten text recognition and mathematical expression recognition). However, such specialist models usually cannot effectively generalize across different sub-tasks. Recently, generalist models (such as GPT-4V), trained on tremendous data in a unified way, have shown enormous potential in reading text in various scenarios, but with the drawbacks of limited accuracy and low efficiency. In this work, we propose Platypus, a generalized specialist model for text reading. Specifically, Platypus combines the best of both worlds: being able recognize text of various forms with a single unified architecture, while achieving excellent accuracy and high efficiency. To better exploit the advantage of Platypus, we also construct a text reading dataset (called Worms), the images of which are curated from previous datasets and partially re-labeled. Experiments on standard benchmarks demonstrate the effectiveness and superiority of the proposed Platypus model. Model and data will be made publicly available at https://github.com/AlibabaResearch/AdvancedLiterateMachinery/tree/main/OCR/Platypus.

Contrastive Language-Image Pre-training (CLIP) has been the cornerstone for zero-shot classification, text-image retrieval, and text-image generation by aligning image and text modalities. Despite its widespread adoption, a significant limitation of CLIP lies in the inadequate length of text input. The length of the text token is restricted to 77, and an empirical study shows the actual effective length is even less than 20. This prevents CLIP from handling detailed descriptions, limiting its applications for image retrieval and text-to-image generation with extensive prerequisites. To this end, we propose Long-CLIP as a plug-and-play alternative to CLIP that supports long-text input, retains or even surpasses its zero-shot generalizability, and aligns the CLIP latent space, making it readily replace CLIP without any further adaptation in downstream frameworks. Nevertheless, achieving this goal is far from straightforward, as simplistic fine-tuning can result in a significant degradation of CLIP's performance. Moreover, substituting the text encoder with a language model supporting longer contexts necessitates pretraining with vast amounts of data, incurring significant expenses. Accordingly, Long-CLIP introduces an efficient finetuning solution on CLIP with two novel strategies designed to maintain the original capabilities, including (1) a knowledge-preserved stretching of positional embedding and (2) a primary component matching of CLIP …

Recently, zero-shot image captioning has gained increasing attention, where only text data is available for training. The remarkable progress in text-to-image diffusion model presents the potential to resolve this task by employing synthetic image-caption pairs generated by this pre-trained prior. Nonetheless, the defective details in the salient regions of the synthetic images introduce semantic misalignment between the synthetic image and text, leading to compromised results. To address this challenge, we propose a novel Patch-wise Cross-modal feature Mix-up (PCM) mechanism to adaptively mitigate the unfaithful contents in a fine-grained manner during training, which can be integrated into most of encoder-decoder frameworks, introducing our PCM-Net. Specifically, for each input image, salient visual concepts in the image are first detected considering the image-text similarity in CLIP space. Next, the patch-wise visual features of the input image are selectively fused with the textual features of the salient visual concepts, leading to a mixed-up feature map with less defective content. Finally, a visual-semantic encoder is exploited to refine the derived feature map, which is further incorporated into the sentence decoder for caption generation. Additionally, to facilitate the model training with synthetic data, a novel CLIP-weighted cross-entropy loss is devised to prioritize the high-quality image-text pairs …

In this paper we propose a novel modification of Contrastive Language-Image Pre-Training (CLIP) guidance for the task of unsupervised backlit image enhancement. Our work builds on the state-of-the-art CLIP-LIT approach, which learns a prompt pair by constraining the text-image similarity between a prompt (negative/positive sample) and a corresponding image (backlit image/well-lit image) in the CLIP embedding space. Learned prompts then guide an image enhancement network. Based on the CLIP-LIT framework, we propose two novel methods for CLIP guidance. First, we show that instead of tuning prompts in the space of text embeddings, it is possible to directly tune their embeddings in the latent space without any loss in quality. This accelerates training and potentially enables the use of additional encoders that do not have a text encoder. Second, we propose a novel approach that does not require any prompt tuning. Instead, based on CLIP embeddings of backlit and well-lit images from training data, we compute the residual vector in the embedding space as a simple difference between the mean embeddings of the well-lit and backlit images. This vector then guides the enhancement network during training, pushing a backlit image towards the space of well-lit images. This approach further dramatically reduces …

We present a unified, promptable model capable of simultaneously segmenting, recognizing, and captioning anything. Unlike SAM, we aim to build a versatile region representation in the wild via visual prompting. To achieve this, we train a generalizable model with massive segmentation masks, e.g., SA-1B masks, and semantic priors from a pre-trained CLIP model with 5 billion parameters. Specifically, we construct a promptable image decoder by adding a semantic token to each mask token. The semantic token is responsible for learning the semantic priors in a predefined concept space. Through joint optimization of segmentation on mask tokens and concept prediction on semantic tokens, our model exhibits strong regional recognition and localization capabilities. For example, an additional 38M-parameter causal text decoder trained from scratch sets a new record with a CIDEr score of 164.7 on the Visual Genome region captioning task. We believe this model can be a versatile region-level image tokenizer, capable of encoding general-purpose region context for a broad range of visual perception tasks.

Learning visual representation with image-text datasets attracts a lot of attention in recent years. Existing approaches primarily rely on cross-modality supervision, and incorporate intra-modality supervision if necessary. They overlook the potential benefits of modality-fused supervision. Since modality-fused representation augments the image representation with textual information, we conjecture it is more discriminative and potential to be a strong teacher for visual representation learning. In this paper, we validate this hypothesis by experiments and propose a novel method FuseTeacher that learns visual representation by modality-fused supervision. Specifically, we introduce a fusion encoder that encodes image and text into a fusion representation. This representation can be utilized to supervise the visual representation learning in two distillation ways: (i) Classification Distillation: we cluster image-text pairs into K clusters using the fusion representation and assign each pair a soft cluster assignment, which is served as a pseudo classification label for supervising the image encoder. (ii) Retrieval Distillation: we calculate the similarities between the fusion representation and all text representations in the same batch. By using the similarity distribution as pseudo retrieval similarity between the corresponding image and all texts, we can prevent one-to-one contrastive learning from separating relevant but unpaired pairs. The FuseTeacher is compatible …

Deep-learning based gaze estimation methods suffer from sever performance degradation in cross-domain settings. One of the primary reason is that the gaze estimation model is confounded by gaze-irrelevant factor during estimation, such as identity and illumination. In this paper, we propose to tackle this problem by causal intervention, an analytical tool that alleviates the impact of confounding factors by using intervening the distribution of confounding factors. Concretely, we propose the Feature-Separation-based Causal Intervention (FSCI) framework for generalizable gaze estimation. The FSCI framework first separates gaze features from gaze-irrelevant features. To alleviate the impact of gaze-irrelevant factors during training, the FSCI framework further implements causal intervention by averaging gaze-irrelevant features using the proposed Dynamic Confounder Bank strategy. Experiments show that the proposed FSCI framework outperforms SOTA gaze estimation methods in varies cross-domain settings, improving the cross-domain accuracy of the baseline up to 36.2% without touching target domain data.

Prompt learning has been widely adopted to efficiently adapt vision-language models (VLMs), e.g. CLIP, for few-shot image classification. Despite their success, most prompt learning methods trade-off between classification accuracy and robustness, e.g. in domain generalization or out-of-distribution (OOD) detection. In this work, we introduce Global-Local Prompts (GalLoP), a new prompt learning method that learns multiple diverse prompts leveraging both global and local visual features. The training of the local prompts relies on local features with an enhanced vision-text alignment. To focus only on pertinent features, this local alignment is coupled with a sparsity strategy in the selection of the local features. We enforce diversity on the set of prompts using a new prompt dropout'' technique and a multiscale strategy on the local prompts. GalLoP outperforms previous prompt learning methods on accuracy on eleven datasets in different few shots settings and with various backbones. Furthermore, GalLoP shows strong robustness performances in both domain generalization and OOD detection, even outperforming dedicated OOD detection methods. Code and instructions to reproduce our results will be open-sourced.

Exploiting the foundation models (\eg, CLIP) to build a versatile keypoint detector has gained increasing attention. Most existing models accept either the text prompt (\eg, the nose of a cat''), or the visual prompt (\eg, support image with keypoint annotations), to detect the corresponding keypoints in query image, thereby, exhibiting either zero-shot or few-shot detection ability. However, the research on taking multimodal prompt is still underexplored, and the prompt diversity in semantics and language is far from opened. For example, how to handle unseen text prompts for novel keypoint detection and the diverse text prompts likeCan you detect the nose and ears of a cat?'' In this work, we open the prompt diversity from three aspects: modality, semantics (seen \vs unseen), and language, to enable a more generalized zero- and few-shot keypoint detection (Z-FSKD). We propose a novel OpenKD model which leverages multimodal prototype set to support both visual and textual prompting. Further, to infer the keypoint location of unseen texts, we add the auxiliary keypoints and texts interpolated from visual and textual domains into training, which improves the spatial reasoning of our model and significantly enhances zero-shot novel keypoint detection. We also found large language model (LLM) is …

The depth/thermal information is beneficial for detecting salient object with conventional RGB images. However, in dual-modal salient object detection (SOD) model, the robustness against noisy inputs and modality missing is crucial but rarely studied. To tackle this problem, we introduce \textbf{Co}nditional Dropout and \textbf{LA}nguage-driven(\textbf{CoLA}) framework comprising two core components. 1) Language-driven Quality Assessment (LQA): Leveraging a pretrained vision-language model with a prompt learner, the LQA recalibrates image contributions without requiring additional quality annotations. This approach effectively mitigates the impact of noisy inputs. 2) Conditional Dropout (CD): A learning method to strengthen the model's adaptability in scenarios with missing modalities, while preserving its performance under complete modalities. The CD serves as a plug-in training scheme that treats modality-missing as conditions, strengthening the overall robustness of various dual-modal SOD models. Extensive experiments demonstrate that the proposed method outperforms state-of-the-art dual-modal SOD models, under both modality-complete and modality-missing conditions.

Replicating the innate human ability to detect all objects based on free-form texts at any granularity remains a formidable challenge for Large Vision Language Models (LVLMs). Current LVLMs are predominantly constrained to locate a single, pre-existing object. This limitation leads to a compromise in model design, necessitating the introduction of visual expert models or the customized head structures. Beyond these constraints, our research uncovers LVLMs' capability for basic object perception, allowing them to accurately identify and locate objects of interest. Building on this insight, we introduce a novel Language-prompted Localization Dataset to fully unleash the capabilities of LVLMs in fine-grained object perception and precise location awareness. More importantly, we present Griffon, a purely LVLM-based baseline, which does not introduce any special tokens, expert models, or additional detection modules. It simply maintains a consistent structure with popular LVLMs by unifying data formats across various localization-related scenarios and is trained end-to-end through a well-designed pipeline. Comprehensive experiments demonstrate that Griffon not only achieves state-of-the-art performance on the fine-grained RefCOCO series and Flickrs30K Entities but also approaches the capabilities of the expert model Faster RCNN on the detection benchmark MSCOCO. Dataset, codes and models will be released.

Out-of-distribution (OOD) object detection is a challenging task due to the absence of open-set OOD data. Inspired by recent advancements in text-to-image generative models, such as Stable Diffusion, we study the potential of generative models trained on large-scale open-set data to synthesize OOD samples, thereby enhancing OOD object detection. We introduce SyncOOD, a simple data curation method that capitalizes on the capabilities of large foundational models to automatically extract meaningful OOD data from text-to-image generative models. This offers the model access to open-world knowledge encapsulated within off-the-shelf foundational models. The synthetic OOD samples are then employed to augment the training of a lightweight, plug-and-play OOD detector, thus effectively optimizing the in-distribution (ID)/OOD decision boundaries. Extensive experiments across multiple benchmarks demonstrate that SyncOOD significantly outperforms existing methods, establishing new state-of-the-art performance with minimal synthetic data usage. The code and data will be publicly available.

Perceiving the world as 3D occupancy supports embodied agents to avoid collision with any type of obstacle. While open-vocabulary image understanding has prospered recently, how to bind the predicted 3D occupancy grids with open-world semantics still remains under-explored due to limited open-world annotations. Hence, instead of building our model from scratch, we try to blend 2D foundation models, specifically a depth model MiDaS and a semantic model CLIP, to lift the semantics to 3D space, thus fulfilling 3D occupancy. However, building upon these foundation models is not trivial. First, the MiDaS faces the depth ambiguity problem, i.e., it only produces coarse relative depth and fails to estimate the one-hot bin depth. Second, the CLIP image features lack high-resolution pixel-level information, which limits the 3D occupancy accuracy. Third, open vocabulary is often trapped by the long-tail problem. To address these issues, we propose VEON for Vocabulary-Enhanced Occupancy predictioN by not only assembling but also adapting these foundation models. We first equip MiDaS with a Zoedepth head and low-rank adaptation (LoRA) for relative-metric-bin depth transformation while reserving beneficial depth prior. Then, a lightweight side adaptor network is attached to the CLIP vision encoder to generate high-resolution features for fine-grained 3D occupancy prediction. …

As a core of Vision Transformer (ViT), self-attention has high flexibility in modeling long-range dependencies because every query attends to all spatial locations. Although ViT achieves promising performance in visual tasks, self-attention's complexity is quadratic with token lengths. This leads to challenging problems when transferring ViT models to dense prediction tasks that require high input resolutions. Previous arts have tried to solve this problem by introducing sparse attention such as spatial reduction attention, and window attention. One common point of these methods is that all image/window tokens are joined during computing attention weights. In this paper, we find out that there exist high similarities between attention weights and incur computation redundancy. To address this issue, this paper proposes novel attention, called partial attention, that learns spatial interactions more efficiently, by reducing redundant information in attention maps. Each query in our attention only interacts with a small set of relevant tokens. Based on partial attention, we design an efficient and general vision Transformer, named PartialFormer, that attains good trade-offs between accuracy and computational costs across vision tasks. For example, on ImageNet-1K, PartialFormer-B3 outperforms Swin-T by 1.7% Top-1 accuracy while saving 25% GFLOPs, and Focal-T by 0.8% while saving 30% GFLOPs.

Referring Expression Segmentation (RES) aims to provide a segmentation mask of the target object in an image referred to by the text (i.e., referring expression). Existing methods require large-scale mask annotations. Moreover, such approaches do not generalize well to unseen/zero-shot scenarios. To address the aforementioned issues, we propose a weakly-supervised bootstrapping architecture for RES with several new algorithmic innovations. To the best of our knowledge, ours is the first approach that considers only a fraction of both mask and box annotations (shown in Figure 1 and Table 1) for training. To enable principled training of models in such low-annotation settings, improve image-text region-level alignment, and further enhance spatial localization of the target object in the image, we propose Cross-modal Fusion with Attention Consistency module. For automatic pseudo-labeling of unlabeled samples, we introduce a novel Mask Validity Filtering routine based on a spatially aware zero-shot proposal scoring approach. Extensive experiments show that with just 30% annotations, our model SAFARI achieves 59.31 and 48.26 mIoUs as compared to 58.93 and 48.19 mIoUs obtained by the fully-supervised SOTA method SeqTR respectively on RefCOCO+@testA and RefCOCO+testB datasets. SAFARI also outperforms SeqTR by 11.7% (on RefCOCO+testA) and 19.6% (on RefCOCO+testB) in a fully-supervised setting and …

Referring Image Segmentation (RIS) leveraging transformers has achieved great success on the interpretation of complex visual-language tasks. However, the quadratic computation cost makes it difficult in capturing long-range visual-language dependencies, which is particularly important for the context of large-size images with long textual descriptions. Fortunately, Mamba addresses this with efficient linear complexity in processing. However, directly applying Mamba to multi-modal interactions presents challenges, primarily due to inadequate channel interactions for the effective fusion of multi-modal data. In this paper, we propose \methodname, a novel RIS architecture that integrates the efficiency of Mamba with a multi-modal Mamba Twister block. The Mamba Twister explicitly models image-text interaction, and fuses textual and visual features through its unique channel and spatial twisting mechanism. We achieve state-of-the-art on all three benchmarks. Moreover, we conduct thorough analyses of \methodname and discuss other fusion designs using Mamba. These provide valuable perspectives for future research. The code will be released upon publication.

Existing scene text removal (STR) task suffers from insufficient training data due to the expensive pixel-level labeling. In this paper, we aim to address this issue by introducing a Text-aware Masked Image Modeling algorithm (TMIM), which can pretrain STR models with low-cost text detection labels (e.g., text bounding box). Different from previous pretraining methods that use indirect auxiliary tasks only to enhance the implicit feature extraction ability, our TMIM first enables the STR task to be directly trained in a weakly supervised manner, which explores the STR knowledge explicitly and efficiently. In TMIM, first, a Background Modeling stream is built to learn background generation rules by recovering the masked non-text region. Meanwhile, it provides pseudo STR labels on the masked text region. Second, a Text Erasing stream is proposed to learn from the pseudo labels and equip the model with end-to-end STR ability. Benefiting from the two collaborative streams, our STR model can achieve impressive performance only with the public text detection datasets, which greatly alleviates the limitation of the high-cost STR labels. Experiments demonstrate that our method outperforms other pretrain methods and achieves state-of-the-art performance (37.35 PSNR on SCUT-EnsText). Code will be available.

This paper introduces ProLab, a novel approach using property-level label space for creating strong interpretable segmentation models. Instead of relying solely on category-specific annotations, ProLab uses descriptive properties grounded in common sense knowledge for supervising segmentation models. It is based on two core designs. First, we employ Large Language Models (LLMs) and carefully crafted prompts to generate descriptions of all involved categories that carry meaningful common sense knowledge and follow a structured format. Second, we introduce a description embedding model preserving semantic correlation across descriptions and then cluster them into a set of descriptive properties (e.g., 256) using K-Means. These properties are based on interpretable common sense knowledge consistent with theories of human recognition. We empirically show that our approach makes segmentation models perform stronger on five classic benchmarks (e.g., ADE20K, COCO-Stuff, Pascal Context, Cityscapes, and BDD). Our method also shows better scalability with extended training steps than category-level supervision. Our interpretable segmentation framework also emerges with the generalization ability to segment out-of-domain or unknown categories using only in-domain descriptive properties. Codes will be publicly available.

Curvilinear object segmentation plays a crucial role across various applications, yet datasets in this domain often suffer from small scale due to the high costs associated with data acquisition and annotation. To address these challenges, this paper introduces a novel approach for expanding curvilinear object segmentation datasets, focusing on enhancing the informativeness of generated data and the congruence between semantic maps and generated images. Our method enriches synthetic data informativeness by generating curvilinear objects through their multiple textual features. By combining textual features from each sample in original dataset, we obtain synthetic images that beyond the original dataset's distribution. This initiative necessitated the creation of the Curvilinear Object Segmentation based on Text Generation (COSTG) dataset. Designed to surpass the limitations of conventional datasets, COSTG incorporates not only standard semantic maps but also some textual descriptions of curvilinear object features. To ensure the congruence between synthetic semantic maps and images, we adapted ControlNet with Spatially-Adaptive Normalization (SPADE), allowing it to preserve semantic information that would typically be washed away in normalization layers. This modification facilitates more accurate semantic image synthesis. Experimental results demonstrate the efficacy of our approach across three types of curvilinear objects (angiography, crack and retina) and five public …

TBA

Large-scale vision foundation models such as Segment Anything (SAM) demonstrate impressive performance in zero-shot image segmentation at multiple levels of granularity. However, these zero-shot predictions are rarely 3D consistent. As the camera viewpoint changes in a scene, so do the segmentation predictions, as well as the characterizations of “coarse” or “fine” granularity. In this work, we address the challenging task of lifting multi-granular and view-inconsistent image segmentations into a hierarchical and 3D-consistent representation. We learn a novel feature field within a Neural Radiance Field (NeRF) representing a 3D scene, whose segmentation structure can be revealed at different scales by simply using different thresholds on feature distance. Our key idea is to learn an ultrametric feature space, which unlike a Euclidean space, exhibits transitivity in distance-based grouping, naturally leading to a hierarchical clustering. Put together, our method takes view inconsistent multi-granularity 2D segmentations as input and produces a hierarchy of 3D-consistent segmentations as output. We evaluate our method and several baselines on a synthetic dataset with multi-view images and multi granular segmentation, showcasing improved accuracy and viewpoint-consistency. We additionally provide qualitative examples of our model’s 3D hierarchical segmentations in real world scenes.

Weakly Supervised Object Localization (WSOL), which aims to localize objects by only using image-level labels, has attracted much attention because of its low annotation cost in real applications. Current studies focus on the Class Activation Map (CAM) of CNN and the self-attention map of transformer to identify the region of objects. However, both CAM and self-attention maps can not learn pixel-level fine-grained information on the foreground objects, which hinders the further advance of WSOL. To address this problem, we initiatively leverage the capability of zero-shot generalization and fine-grained segmentation in Segment Anything Model (SAM) to boost the activation of integral object regions. Further, to alleviate the semantic ambiguity issue accrued in single point prompt-based SAM, we propose an innovative mask prompt to SAM (Pro2SAM) network with grid points for WSOL task. First, we devise a Global Token Transformer (GTFormer) to generate a coarse-grained foreground map as a flexible mask prompt, where the GTFormer jointly embeds patch tokens and novel global tokens to learn foreground semantics. Secondly, we deliver grid points as dense prompts into SAM to maximize the probability of foreground mask, which avoids the lack of objects caused by a single point/box prompt. Finally, we propose a pixel-level similarity …

Semantic segmentation is an important task for many applications but it is still quite challenging to achieve advanced performance with limited computational costs. In this paper, we present CGRSeg, an efficient yet competitive segmentation framework based on context-guided spatial feature reconstruction. In it, a Rectangular Self-Calibration Module is carefully designed for spatial feature reconstruction and pyramid context extraction. It captures the global context in both horizontal and vertical directions and gets the axial global context to explicitly model rectangular key areas. A shape self-calibration function is designed to make the key areas more close to the foreground object. Besides, a lightweight Dynamic Prototype Guided head is proposed to improve the classification of foreground objects by explicit class embedding. Our CGRSeg is extensively evaluated on ADE20K, COCO-Stuff, and Pascal Context benchmarks, and achieves state-of-the-art semantic performance. Specifically, it achieves 43.6% mIoU on ADE20K with only 4.0 GFLOPs, which is 0.9% and 2.5% mIoU better than SeaFormer and SegNeXt but with about 38.0% fewer GFLOPs.

We present PartGLEE, a part-level foundation model for locating and identifying both objects and parts in images. Through a unified framework, PartGLEE accomplishes detection, segmentation, and grounding of instances at any granularity in the open world scenario. Specifically, we propose a Q-Former to construct the hierarchical relationship between objects and parts, parsing every object into corresponding semantic parts. By incorporating a large amount of object-level data, the hierarchical relationships can be extended, enabling PartGLEE to recognize a rich variety of parts. We conduct comprehensive empirical studies to validate the effectiveness of our method, PartGLEE achieves the state-of-the-art performance across various part-level tasks and maintain comparable results on object-level tasks. Our further analysis indicates that the hierarchical cognitive ability of PartGLEE is able to facilitate a detailed comprehension in images for mLLMs. Code will be released.

Current 3D scene segmentation methods are heavily dependent on manually annotated 3D training datasets. Such manual annotations are labor-intensive, and often lack fine-grained details. Furthermore, models trained on this data typically struggle to recognize object classes beyond the annotated training classes, i.e., they do not generalize well to unseen domains and require additional domain-specific annotations. In contrast, recent 2D foundation models have demonstrated strong generalization and impressive zero-shot abilities, inspiring us to incorporate these characteristics from 2D models into 3D models. Therefore, we explore the use of image segmentation foundation models to automatically generate high-quality training labels for 3D segmentation models. The resulting model, Segment3D, generalizes significantly better than the models trained on costly manual 3D labels and enables easily adding new training data to further boost the segmentation performance.

Open-world 3D instance segmentation is a recently introduced problem with diverse applications, notably in continually learning embodied agents. This task involves segmenting unknown instances, and learning new instances when their labels are introduced. However, prior research in the open-world domain has traditionally addressed the two sub-problems, namely continual learning and unknown object identification, separately. This approach has resulted in limited performance on unknown instances and cannot effectively mitigate catastrophic forgetting. Additionally, these methods bypass the utilization of the information stored in the previous version of the continual learning model, instead relying on a dedicated memory to store historical data samples, which inevitably leads to an expansion of the memory budget. In this paper, we argue that continual learning and unknown class identification should be tackled in conjunction. Therefore, we propose a new exemplar-free approach for 3D continual learning and the discovery of unknown classes through self-distillation. Our approach leverages the pseudo-labels generated by the model from the preceding task to improve the unknown predictions during training while simultaneously mitigating catastrophic forgetting. By integrating these pseudo-labels into the continual learning process, we achieve enhanced performance in handling unknown classes. We validate the efficacy of the proposed approach via comprehensive experiments on …

Class activation maps (CAMs) are commonly employed in weakly supervised semantic segmentation (WSSS) to produce pseudo-labels. Due to incomplete or excessive class activation, existing studies often resort to offline CAM refinement, introducing additional stages or proposing offline modules. This can cause optimization difficulties for single-stage methods and limit generalizability. In this study, we aim to reduce the observed CAM inconsistency and error to mitigate reliance on refinement processes. We propose an end-to-end WSSS model incorporating guided CAMs, wherein our segmentation model is trained while concurrently optimizing CAMs online. Our method, Co-training with Swapping Assignments (CoSA), leverages a dual-stream framework, where one sub-network learns from the swapped assignments generated by the other. We introduce three techniques in this framework: i) soft perplexity-based regularization to penalize uncertain regions; ii) a threshold-searching approach to dynamically revise the confidence threshold; and iii) contrastive separation to address the coexistence problem. CoSA demonstrates exceptional performance, achieving mIoU of 76.2% and 51.0% on VOC and COCO validation datasets, respectively, surpassing existing baselines by a substantial margin. Notably, CoSA is the first single-stage approach to outperform all existing multi-stage methods including those with additional supervision. Source code will be publicly available.

Incorporating pixel contextual information is critical for accurate segmentation. In this paper, we show that an effective way to incorporate contextual information is through a patch-based classifier. This patch classifier is trained to identify classes present within an image region, which facilitates the elimination of distractors and enhances the classification of small object segments. Specifically, we introduce \textbf{Multi-scale Patch-based Multi-label Classifier} (MPMC), a novel plug-in module designed for existing semi-supervised segmentation (SSS) frameworks. MPMC offers patch-level supervision, enabling the discrimination of pixel regions of different classes within a patch. Furthermore, MPMC learns an adaptive pseudo-label weight, using patch-level classification to alleviate the impact of the teacher’s noisy pseudo-label supervision on the student. This lightweight module can be integrated into any SSS framework, significantly enhancing their performance. We demonstrate the efficacy of our proposed MPMC by integrating it into four SSS methodologies and improving them across two natural image and one medical segmentation dataset, notably improving the segmentation results of the baselines across all the three datasets.

3D segmentation is a core problem in computer vision and, similarly to many other dense prediction tasks, it requires large amounts of annotated data for adequate training. However, densely labeling 3D point clouds to employ fully-supervised training remains too labor intensive and expensive. Semi-supervised training provides a more practical alternative, where only a small set of labeled data is given, accompanied by a larger unlabeled set. This area thus studies the effective use of unlabeled data to reduce the performance gap that arises due to the lack of annotations. In this work, inspired by Bayesian deep learning, we first propose a Bayesian self-training framework for semi-supervised 3D semantic segmentation. Employing stochastic inference, we generate an initial set of pseudo-labels and then filter these based on estimated point-wise uncertainty. By constructing a heuristic n-partite matching algorithm, we extend the method to semi-supervised 3D instance segmentation, and finally, with the same building blocks, to dense 3D visual grounding. We demonstrate state-of-the-art results for our semi-supervised method on SemanticKITTI and ScribbleKITTI for 3D semantic segmentation and on ScanNet and S3DIS for 3D instance segmentation. We further achieve substantial improvements in dense 3D visual grounding over supervised-only baselines on ScanRefer.

Point cloud few-shot semantic segmentation (PC-FSS) aims to segment targets of novel categories in a given query point cloud with only a few annotated support samples. The current top-performing prototypical learning methods employ prototypes originating from support samples to direct the classification of query points. However, the inherent fragility of point-level matching and the prevalent intra-class diversity pose great challenges to this cross-instance matching paradigm, leading to erroneous background activations or incomplete target excavation. In this work, we propose a simple yet effective framework in the spirit of Decoupled Localization and Expansion (DLE). The proposed DLE, including a structural localization module (SLM) and a self-expansion module (SEM), enjoys several merits. First, structural information is injected into the matching process through the agent-level correlation in SLM, and the confident target region can thus be precisely located. Second, more reliable intra-object similarity is harnessed in SEM to derive the complete target, and the conservative expansion strategy is introduced to reasonably constrain the expansion. Extensive experiments on two challenging benchmarks under different settings demonstrate that DLE outperforms previous state-of-the-art approaches by large margins.

Large-scale 3D bounding box annotation is crucial for LiDAR object detection but comes at a high cost. Semi-supervised object detection (SSOD) offers promising solutions to leverage unannotated data, but the predominant pseudo-labeling approach requires careful hyperparameter tuning for training on noisy teacher labels. In this work, we propose a Cross-Scan Object Transfer (CSOT) paradigm for LiDAR SSOD. Central to our approach is Hotspot Network, a transformer-based network that predicts possible placement locations for annotated objects in unannotated scans and assigns scores to each location. By leveraging these contextual consistent location predictions, CSOT successfully enables object copy-paste in LiDAR SSOD for the first time. To train object detectors on partially annotated scans generated by CSOT, we adopt a Spatial-Aware classification loss throughout our partial supervision to handle false negative issues caused by treating all unlabeled objects as background. We conduct extensive experiments to verify the efficacy and generality of our method. Compared to other state-of-the-art label-efficient methods used in LiDAR detection, our approach requires the least amount of annotation while achieves the best detectior. Using only 1% of the labeled data on the Waymo dataset, our semi-supervised detector achieves performance on par with the fully supervised baseline. Similarly, on the nuScenes …

The evolution of 3D object detection hinges not only on advanced models but also on effective and efficient annotation strategies. Despite this progress, the labor-intensive nature of 3D object annotation remains a bottleneck, hindering further development in the field. This paper introduces a novel approach, incorporated with prompt in 2D, detect in 3D'' anddetect in 3D, refine in 3D'' strategies, to 3D object annotation: multi-modal interactive 3D object detection. Firstly, by allowing users to engage with simpler 2D interaction prompts (e.g., clicks or boxes on a camera image or a bird's eye view), we bridge the complexity gap between 2D and 3D spaces, reimagining the annotation workflow. Besides, Our framework also supports flexible iterative refinement to the initial 3D annotations, further assisting annotators in achieving satisfying results. Evaluation on the nuScenes dataset demonstrates the effectiveness of our method. And thanks to the prompt-driven and interactive designs, our approach also exhibits outstanding performance in open-set scenarios. This work not only offers a potential solution to the 3D object annotation problem but also paves the way for further innovations in the 3D object detection community.

Most Camouflaged Object Detection (COD) methods heavily rely on mask annotations, which are time-consuming and labor-intensive to acquire. Existing weakly-supervised COD approaches exhibit significantly inferior performance compared to fully-supervised methods and struggle to simultaneously support all the existing types of camouflaged object labels, including scribbles, bounding boxes, and points. Even for Segment Anything Model (SAM), it is still problematic to handle the weakly-supervised COD and it typically encounters challenges of prompt compatibility of the scribble labels, extreme response, semantically erroneous response, and unstable feature representations, producing unsatisfactory results in camouflaged scenes. To mitigate these issues, we propose a unified COD framework in this paper, termed SAM-COD, which is capable of supporting arbitrary weakly-supervised labels. Our SAM-COD employs a prompt adapter to handle scribbles as prompts based on SAM. Meanwhile, we introduce response filter and semantic matcher modules to improve the quality of the masks obtained by SAM under COD prompts. To alleviate the negative impacts of inaccurate mask predictions, a new strategy of prompt-adaptive knowledge distillation is utilized to ensure a reliable feature representation. To validate the effectiveness of our approach, we have conducted extensive empirical experiments on three mainstream COD benchmarks. The results demonstrate the superiority of our method …

Modern pre-trained architectures struggle to retain previous information while undergoing continuous fine-tuning on new tasks. Despite notable progress in continual classification, systems designed for complex vision tasks such as detection or segmentation still struggle to attain satisfactory performance. In this work, we introduce a memory-based detection transformer architecture to adapt a pre-trained DETR-style detector to new tasks while preserving knowledge from previous tasks. We propose a novel localized query function for efficient information retrieval from memory units, aiming to minimize forgetting. Furthermore, we identify a fundamental challenge in continual detection referred to as {\em background relegation}. This arises when object categories from earlier tasks reappear in future tasks, potentially without labels, leading them to be implicitly treated as background. This is an inevitable issue in continual detection or segmentation. The introduced continual optimization technique effectively tackles this challenge. Finally, we assess the performance of our proposed system on continual detection benchmarks and demonstrate that our approach surpasses the performance of existing state-of-the-art resulting in 5-7\% improvements on MSCOCO and PASCAL-VOC on the task of continual detection.

The Common Objects in Context (COCO) dataset has been instrumental in benchmarking object detectors over the past decade. Like every dataset, COCO contains subtle errors and imperfections stemming from its annotation procedure. With the advent of high-performing models, we ask whether these errors of COCO are hindering its utility in reliably benchmarking further progress. In search for an answer, we inspect thousands of masks from COCO (2017 version) and uncover different types of errors such as imprecise mask boundaries, non-exhaustively annotated instances, and mislabeled masks. Due to the prevalence of COCO, we choose to correct these errors to maintain continuity with prior research. We develop COCO-ReM (Refined Masks), a cleaner set of annotations with visibly better mask quality than COCO-2017. We evaluate fifty object detectors and find that models that predict visually sharper masks score higher on COCO-ReM, affirming that they were being incorrectly penalized due to errors in COCO-2017. Moreover, our models trained using COCO-ReM converge faster and score higher than their larger variants trained using COCO-2017, highlighting the importance of data quality in improving object detectors. With these findings, we advocate using COCO-ReM for future object detection research. Our dataset is available at https://cocorem.xyz

Camouflaged object detection (COD) has attracted a lot of attention in computer vision. The main challenge lies in the high degree of similarity between camouflaged objects and their surroundings in the spatial domain, making identification difficult. Existing methods attempt to reduce the impact of pixel similarity by maximizing the distinguishing ability of spatial features with complicated design, but often ignore the sensitivity and locality of features in the spatial domain, leading to sub-optimal results. In this paper, we propose a new approach to address this issue by jointly exploring the representation in the frequency and spatial domains, introducing the Frequency-Spatial Entanglement Learning (FSEL) method. This method consists of a series of well-designed Entanglement Transformer Blocks (ETB) for representation learning, a Joint Domain Perception Module (JDPM) for semantic enhancement, and a Dual-domain Reverse Parser (DRF) for feature integration in the frequency and spatial domains. Specifically, the ETB utilizes frequency self-attention (FSA) to effectively characterize the relationship between different frequency bands, while the entanglement feed-forward network (EFFN) facilitates information interaction between features of different domains through entanglement learning. Our extensive experiments demonstrate the superiority of FSEL over 21 state-of-the-art (SOTA) methods, through comprehensive quantitative and qualitative comparisons in three widely-used COD datasets.

Oriented object detection, an emerging task in recent years, aims to identify and locate objects across varied orientations. This requires the detector to accurately capture the orientation information, which varies significantly within and across images. Despite the existing substantial efforts, simultaneously ensuring model effectiveness and parameter efficiency remains challenging in this scenario. In this paper, we propose a lightweight yet effective Group-wise Rotating and Attention (GRA) module to replace the convolution operations in backbone networks for this task. GRA can adaptively capture fine-grained features of objects with diverse orientations, comprising two key components: Group-wise Rotating and Group-wise Attention. Group-wise Rotating firstly divides the convolution kernel into groups, where each group extracts different object features by rotating at a specific angle according to the object orientation. Subsequently, Group-wise Attention is employed to adaptively enhance the object-related regions in the feature. The collaborative effort of these components enables GRA to effectively capture the various orientation information while maintaining parameter efficiency. Extensive experimental results demonstrate the superiority of our method. For example, GRA achieves a new state-of-the-art (SOTA) on the DOTA-v2.0 benchmark, while saving the parameters by nearly 50% compared to the previous SOTA method. Code will be released.

Despite previous DETR-like methods having performed successfully in generic object detection, tiny object detection is still a challenging task for them since the positional information of object queries is not customized for detecting tiny objects, whose scale is extraordinarily smaller than general objects. Also, DETR-like methods using a fixed number of queries make them unsuitable for aerial datasets, which mostly contain tiny objects, and the numbers of instances are imbalanced between different images. Thus, we present a simple yet effective model, DQ-DETR, consisting of three components: categorical counting module, counting-guided feature enhancement, and dynamic query selection to solve the above-mentioned problems. DQ-DETR uses the prediction and density maps from the categorical counting module to dynamically adjust the number and positional information of object queries. Our model DQ-DETR outperforms previous CNN-based and DETR-like methods, achieving state-of-the-art mAP 30.2% on the AI-TOD-V2 dataset, which mostly consists of tiny objects.

This work investigates the problem of instance-level image retrieval with re-ranking with the constraint of memory efficiency, ultimately aiming to limit memory usage to 1KB per image. Departing from the prevalent focus on performance enhancements, this work prioritizes the crucial trade-off between performance and memory requirements. The proposed model employs a transformer-based architecture designed to estimate image-to-image similarity by capturing interactions within and across images based on their local descriptors. A distinctive property of the model is the capability for asymmetric similarity estimation. Database images are represented with a smaller number of descriptors compared to query images, enabling performance improvements without increasing memory consumption. To ensure adaptability across different applications, a universal model is introduced that adjusts to varying descriptor set cardinalities during the testing phase. Results on standard benchmarks demonstrate the superiority of our approach over both hand-crafted and learned models. In particular, compared with current state-of-the-art methods that overlook their memory footprint, our approach not only attains superior performance but does so with a significantly reduced memory footprint. We intend to make our code publicly available.

Semi-supervised medical image segmentation has shown promise in training models with limited labeled data. However, current dominant teacher-student based approaches can suffer from the confirmation bias. To address this challenge, we propose AD-MT, an alternate diverse teaching approach in a teacher-student framework. It involves a single student model and two non-trainable teacher models that are momentum-updated periodically and randomly in an alternate fashion. To mitigate the confirmation bias via the diverse supervision, the core of AD-MT lies in two proposed modules: the Random Periodic Alternate (RPA) Updating Module and the Conflict-Combating Module (CCM). The RPA schedules an alternating diverse updating process with complementary data batches, distinct data augmentation, and random switching periods to encourage diverse reasoning from different teaching perspectives. The CCM employs an entropy-based ensembling strategy to encourage the model to learn from both the consistent and conflicting predictions between the teachers. Experimental results demonstrate the effectiveness and superiority of AD-MT on the 2D and 3D medical segmentation benchmarks across various semi-supervised settings.

Nucleus instance segmentation in histology images is crucial for a broad spectrum of clinical applications. Current dominant algorithms rely on regression of nuclear proxy maps. Distinguishing nucleus instances from the estimated maps requires carefully curated post-processing, which is error-prone and parameter-sensitive. Recently, the Segment Anything Model (SAM) has earned huge attention in medical image segmentation, owing to its impressive generalization ability and promptable property. Nevertheless, its potential on nucleus instance segmentation remains largely underexplored. In this paper, we present a novel prompt-driven framework that consists of a nucleus prompter and SAM for automatic nucleus instance segmentation. Specifically, the prompter is developed to generate a unique point prompt for each nucleus, while SAM is fine-tuned to produce its corresponding mask. Furthermore, we propose to integrate adjacent nuclei as negative prompts to enhance model's capability to identify overlapping nuclei. Without complicated post-processing, our proposed method sets a new state-of-the-art performance on three challenging benchmarks. The source code is available in the supplementary materials.

Multiple instance learning (MIL) has been extensively applied to whole slide histopathology image (WSI) analysis. The existing aggregation strategy in MIL, which primarily relies on the first-order distance (e.g., mean difference) between instances, fails to accurately approximate the true feature distribution of each instance, leading to biased slide-level representations. Moreover, the scarcity of WSI observations easily leads to model overfitting, resulting in unstable testing performance and limited generalizability. To tackle these challenges, we propose a new Bayesian nonparametric framework for multiple instance learning, which adopts a cascade of Dirichlet processes (cDP) to incorporate the instance-to-bag characteristic of the WSIs. We perform feature aggregation based on the latent clusters formed by the Dirichlet process, which incorporates the covariances of the patch features and forms more representative clusters. We then perform bag-level prediction with another Dirichlet process model on the bags, which imposes a natural regularization on learning to prevent overfitting and enhance generalizability. Moreover, as a Bayesian nonparametric method, the cDP model can accurately generate posterior uncertainty, which allows for the detection of outlier samples and tumor localization. Extensive experiments on five WSI benchmarks validate the superior performance of our method, as well as its generalizability and ability to estimate uncertainties. …

Current multi-instance learning algorithms for pathology image analysis often require a substantial number of Whole Slide Images for effective training but exhibit suboptimal performance in scenarios with limited learning data. In clinical settings, restricted access to pathology slides is inevitable due to patient privacy concerns and the prevalence of rare or emerging diseases. The emergence of the Few-shot Weakly Supervised WSI Classification accommodates the significant challenge of the limited slide data and sparse slide-level labels for diagnosis. Prompt learning based on the pre-trained models (e.g., CLIP) appears to be a promising scheme for this setting; however, current research in this area is limited, and existing algorithms often focus solely on patch-level prompts or confine themselves to language prompts. This paper proposes a multi-instance prompt learning framework enhanced with pathology knowledge, i.e., integrating visual and textual prior knowledge into prompts at both patch and slide levels. The training process employs a combination of static and learnable prompts, effectively guiding the activation of pre-trained models and further facilitating the diagnosis of key pathology patterns. Lightweight Messenger and Summary layers are introduced to model relationships between patches and slides within the same patient data. Additionally, alignment-wise contrastive losses ensure the feature-level alignment between …

Due to the common content of anatomy, radiology images with their corresponding reports exhibit high similarity. Such inherent data bias can predispose automatic report generation models to learn entangled and spurious representations resulting in misdiagnostic reports. To tackle these, we propose a novel \textbf{Co}unter\textbf{F}actual \textbf{E}xplanations-based framework (CoFE) for radiology report generation. Counterfactual explanations serve as a potent tool for understanding how decisions made by algorithms can be changed by asking what if'' scenarios. By leveraging this concept, CoFE can learn non-spurious visual representations by contrasting the representations between factual and counterfactual images. Specifically, we derive counterfactual images by swapping a patch between positive and negative samples until a predicted diagnosis shift occurs. Here, positive and negative samples are the most semantically similar but have different diagnosis labels. Additionally, CoFE employs a learnable prompt to efficiently fine-tune the pre-trained large language model, encapsulating both factual and counterfactual content to provide a more generalizable prompt representation. Extensive experiments on two benchmarks demonstrate that leveraging the counterfactual explanations enables CoFE to generate semantically coherent and factually complete reports and outperform in terms of language generation and clinical efficacy metrics.

We introduce eCLIP, an enhanced version of the CLIP model that integrates expert annotations in the form of radiologist eye-gaze heatmaps. It tackles key challenges in contrastive multi-modal medical imaging analysis, notably data scarcity and the modality gap'' -- a significant disparity between image and text embeddings that diminishes the quality of representations and hampers cross-modal interoperability. eCLIP integrates a heatmap processor and leverages mixup augmentation to efficiently utilize the scarce expert annotations, thus boosting the model's learning effectiveness. eCLIP is designed to be generally applicable to any variant of CLIP without requiring any modifications of the core architecture. Through detailed evaluations across several tasks, including zero-shot inference, linear probing, cross-modal retrieval, and Retrieval Augmented Generation (RAG) of radiology reports using a frozen Large Language Model, eCLIP showcases considerable improvements in embedding quality. The outcomes reveal enhanced alignment and uniformity, affirming eCLIP's capability to harness high-quality annotations for enriched multi-modal analysis in the medical imaging domain.

Image generation can solve insufficient labeled data issues in defect detection. Most defect generation methods are only trained on a single product without considering the consistencies among multiple products, leading to poor quality and diversity of generated results. To address these issues, we propose DefectDiffu, a novel text-guided diffusion method to model both intra-product background consistency and inter-product defect consistency across multiple products and modulate the consistency perturbation directions to control product type and defect strength, achieving diversified defect image generation. Firstly, we leverage a text encoder to separately provide consistency prompts for background, defect, and fusion parts of the disentangled integrated architecture, thereby disentangling defects and normal backgrounds. Secondly, we propose the double-free strategy to generate defect images through two-stage perturbation of consistency direction, thereby controlling product type and defect strength by adjusting the perturbation scale. Besides, DefectDiffu can generate defect mask annotations utilizing cross-attention maps from the defect part. Finally, to improve the generation quality of small defects and masks, we propose the adaptive attention-enhance loss to increase the attention to defects. Experimental results demonstrate that DefectDiffu surpasses state-of-the-art methods in terms of generation quality and diversity, thus effectively improving downstream defection performance. Moreover, defect perturbation directions can …

When deploying a semantic segmentation model into the real world, it will inevitably encounter semantic classes that were not seen during training. Therefore, to ensure a safe deployment of such systems, it is crucial to accurately evaluate and improve their anomaly segmentation capabilities. However, acquiring and labelling semantic segmentation data is expensive and unanticipated conditions are long-tail and potentially hazardous. Indeed, existing anomaly segmentation datasets capture a limited number of anomalies, lack realism or have strong domain shifts. In this paper, we propose the Placing Objects in Context (POC) pipeline to realistically add any object into any image via diffusion models. POC can be used to easily extend any dataset with an arbitrary number of objects. In our experiments, we present different anomaly segmentation datasets based on POC-generated data and show that POC can improve the performance of recent state-of-the-art anomaly fine-tuning methods across several standardized benchmarks. POC is also effective for learning new classes. For example, we utilize it to augment Cityscapes samples by incorporating a subset of Pascal classes and demonstrate that models trained on such data achieve comparable performance to the Pascal-trained baseline. This corroborates the low synth2real gap of models trained on POC-generated images.

We are exploring an emerging formulation in anomaly detection (AD) where multiple instances of the same object are produced simultaneously and distinctly to address the limitation that using only a single instance may not effectively capture any underlying defects. More specifically, we concentrate on a specific scenario where each object of interest is linked to seven distinct data views/representations. The first six views involve capturing images with a stationary camera under six different lighting conditions, while the seventh view pertains to the 3D normal information. We refer to our intended task as multi-view anomaly detection. To tackle this problem, our approach involves training a view-invariant ControlNet that can produce consistent feature maps regardless of the data views. This training strategy enables us to mitigate the impact of varying lighting conditions and to fuse information from both the RGB color appearance and the 3D normal geometry effectively. Moreover, as the diffusion process is not deterministic, we utilize the DDIM scheme to improve the applicability of our established memory banks of diffusion-based features for anomaly detection inference. To demonstrate the efficacy of our approach, we present extensive ablation studies and state-of-the-art experimental results on the Eyecandies dataset.

In the field of multi-class anomaly detection, reconstruction-based methods derived from single-class anomaly detection face the well-known challenge of learning shortcuts'', wherein the model fails to learn the patterns of normal samples as it should, opting instead for shortcuts such as identity mapping or artificial noise elimination. Consequently, the model becomes unable to reconstruct genuine anomalies as normal instances, resulting in a failure of anomaly detection. To counter this issue, we present a novel unified feature reconstruction-based anomaly detection framework termed RLR (Reconstruct features from a Learnable Reference representation). Unlike previous methods, RLR utilizes learnable reference representations to compel the model to learn normal feature patterns explicitly, thereby prevents the model from succumbing to thelearning shortcuts'' issue. Additionally, RLR incorporates locality constraints into the learnable reference to facilitate more effective normal pattern capture and utilizes a masked learnable key attention mechanism to enhance robustness. Evaluation of RLR on the 15-category MVTec-AD dataset and the 12-category VisA dataset shows superior performance compared to state-of-the-art methods under the unified setting. The code of RLR will be publicly available.

Video Anomaly Detection (VAD) is crucial for applications such as security surveillance and autonomous driving. However, existing VAD methods provide little rationale behind detection, hindering public trust in real-world deployments. In this paper, we approach VAD with a reasoning framework. Although Large Language Models (LLMs) have shown revolutionary reasoning ability, we find that their direct use falls short of VAD. Specifically, the implicit knowledge pre-trained in LLMs focuses on general context and thus may not apply to every specific real-world VAD scenario, leading to inflexibility and inaccuracy. To address this, we propose AnomalyRuler, a novel rule-based reasoning framework for VAD with LLMs. AnomalyRuler comprises two main stages: induction and deduction. In the induction stage, the LLM is fed with few-shot normal reference samples and then summarizes these normal patterns to induce a set of rules for detecting anomalies. The deduction stage follows the induced rules to spot anomalous frames in test videos. Additionally, we design rule aggregation, perception smoothing, and robust reasoning strategies to further enhance AnomalyRuler's robustness. AnomalyRuler is the first reasoning approach for the one-class VAD task, which requires only few-normal-shot prompting without the need for full-shot training, thereby enabling fast adaption to various VAD scenarios. Comprehensive experiments …

Binary Neural Networks (BNNs) offer a promising avenue toward achieving efficient deep-learning models but are hindered by the inherent challenge of aligning noisy floating-point gradients with binary parameters. To address this, we introduce Diode, a groundbreaking optimizer designed explicitly for BNNs that bridges this gap by utilizing the gradient's sign information in a unique, latent-weight-free approach. By focusing on the gradient sign's lower-order moment estimate for parameter updates, Diode uniformly fine-tunes binary parameters, significantly enhancing model convergence without the dependency on 32-bit latent weights or embedding buffers. This paper showcases Diode's superior performance through comprehensive evaluations on a variety of vision and Natural Language Processing (NLP) tasks. Remarkably, Diode advances the state-of-the-art by increasing BNext-18 Top-1 accuracy on ImageNet ILSVRC2012 by 0.96\% with eightfold fewer training iterations. In the case of ReActNet, Diode not only matches but slightly exceeds previous benchmarks without resorting to complex multi-stage optimization strategies, effectively halving the training duration. Additionally, Diode proves its robust generalization capability on the binary BERT architecture within the GLUE benchmark, outperforming the existing BiT design by 3.3\% without data augmentation and establishing a new SOTA accuracy of 78.8\% with augmentation. The implementation of Diode will be made available to the public …

In this paper, we introduce Saliency-Based Adaptive Masking (SBAM), a novel and cost-effective approach that significantly enhances the pre-training performance of Masked Image Modeling (MIM) approaches by prioritizing token salience. Our method provides robustness against variations in masking ratios, effectively mitigating the performance instability issues common in existing methods. This relaxes the sensitivity of MIM-based pre-training to masking ratios, which in turn allows us to propose an adaptive strategy for `tailored' masking ratios for each data sample, which no existing method can provide. Toward this goal, we propose an Adaptive Masking Ratio (AMR) strategy that dynamically adjusts the proportion of masking for the unique content of each image based on token salience. We show that our method significantly improves over the state-of-the-art in mask-based pre-training on the ImageNet-1K dataset.

Multi-head self-attention (MSA) is a key component of Vision Transformers (ViTs), which have achieved great success in various vision tasks. However, their high computational cost and memory footprint hinder their deployment on resource-constrained devices. Conventional pruning approaches can only compress and accelerate the MSA module using head pruning, although the head is not an atomic unit. To address this issue, we propose a novel graph-aware neuron-level pruning method, Structured Neuron-level Pruning (SNP). SNP prunes neurons with less informative attention scores and eliminates redundancy among heads. Specifically, it prunes graphically connected query and key layers having the least informative attention scores while preserving the overall attention scores. Value layers, which can be pruned independently, are pruned to eliminate inter-head redundancy. Our proposed method effectively compresses and accelerates Transformer-based models for both edge devices and server processors. For instance, the DeiT-Small with SNP runs 3.1 times faster than the original model and achieves performance that is 21.94\% faster and 1.12\% higher than the DeiT-Tiny. Additionally, SNP accelerates the efficiently designed Transformer model, EfficientFormer, by 1.74 times on the Jetson Nano with acceptable performance degradation.

This paper introduces TinySaver, an early-exit-like dynamic model compression approach which employs tiny models to substitute large models adaptively. Distinct from traditional compression techniques, dynamic methods like TinySaver can leverage the difficulty differences to allow certain inputs to complete their inference processes early, thereby conserving computational resources. Most existing early exit designs are implemented by attaching additional network branches to the model's backbone. Our study, however, reveals that completely independent tiny models can replace a substantial portion of the larger models' job with minimal impact on performance. Employing them as the first exit can remarkably enhance computational efficiency. By searching and employing the most appropriate tiny model as the computational saver for a given large model, the proposed approaches work as a novel and generic method to model compression. This finding will help the research community in exploring new compression methods to address the escalating computational demands posed by rapidly evolving AI models. Our evaluation of this approach in ImageNet-1k classification demonstrates its potential to reduce the number of compute operations by up to 90\%, with only negligible losses in performance, across various modern vision models. The code of this work will be available.

Token compression expedites the training and inference of Vision Transformers (ViTs) by reducing the number of the redundant tokens, e.g., pruning inattentive tokens or merging similar tokens. However, when applied to downstream tasks, these approaches suffer from significant performance drop when the compression degrees are mismatched between training and inference stages, which limits the application of token compression on off-the-shelf trained models. In this paper, we propose a model arithmetic framework to decouple the compression degrees between the two stages. In advance, we additionally perform a fast parameter-efficient self-distillation stage on the pre-trained model to obtain a small plugin, called Token Compensator (ToCom), which describes the gap between models across different compression degrees. During inference, ToCom can be directly inserted into any downstream off-the-shelf models with any mismatched training and inference compression degrees to acquire universal performance improvements without further training. Experiments on over 20 downstream tasks demonstrate the effectiveness of our framework. On CIFAR100, fine-grained visual classification, and VTAB-1k benchmark, ToCom can yield up to a maximum improvement of 2.3%, 1.5%, and 2.0% in the average performance of DeiT-B, respectively.

Nonlinear activation functions are pivotal to the success of deep neural nets, and choosing the appropriate activation function can significantly affect their performance. Most networks use fixed activation functions (e.g., ReLU, GELU, etc.), which can be sub-optimal as their expressiveness is limited. Furthermore, distinct layers may benefit from diverse activation functions. Consequently, there has been a growing interest in trainable activation functions. In this paper, we introduce DiTAC, a trainable highly-expressive activation function based on an efficient diffeomorphic transformation. Despite introducing only a negligible number of trainable parameters, DiTAC enhances model expressiveness and performance, often yielding substantial improvements. It also outperforms existing activation functions (regardless whether the latter are fixed or trainable) in tasks such as semantic segmentation, image generation, regression problems, and image classification. Our code will be made publicly available upon acceptance.

Traditional deep learning relies on end-to-end backpropagation for training, but it suffers from drawbacks such as high memory consumption and not aligning with biological neural networks. Recent advancements have introduced locally supervised learning, which divides networks into modules with isolated gradients and trains them locally. However, this approach can lead to performance lag due to limited interaction between these modules, and the design of auxiliary networks occupies a certain amount of GPU memory. To overcome these limitations, we propose a novel model called HPFF that performs hierarchical locally supervised learning and patch-level feature computation on the auxiliary networks. Hierarchical Locally Supervised Learning (HiLo) enables the network to learn features at different granularity levels along their respective local paths. Specifically, the network is divided into two-level local modules: independent local modules and cascade local modules. The cascade local modules combine two adjacent independent local modules, incorporating both updates within the modules themselves and information exchange between adjacent modules. Patch Feature Fusion (PFF) reduces GPU memory usage by splitting the input features of the auxiliary networks into patches for computation. By averaging these patch-level features, it enhances the network's ability to focus more on those patterns that are prevalent across multiple patches. …

Transfer learning has long been a key factor in the advancement of many fields including 2D image analysis. Unfortunately, its applicability in 3D data processing has been relatively limited. While several approaches for 3D transfer learning have been proposed in recent literature, with contrastive learning gaining particular prominence, most existing methods in this domain have only been studied and evaluated in limited scenarios. Most importantly, there is currently a lack of principled understanding of both when and why 3D transfer learning methods are applicable. Remarkably, even the applicability of standard supervised pre-training is poorly understood. In this work, we conduct the first in-depth quantitative and qualitative investigation of supervised and contrastive pre-training strategies and their utility in downstream 3D tasks. We demonstrate that layer-wise analysis of learned features provides significant insight into the downstream utility of trained networks. Informed by this analysis, we propose a simple geometric regularization strategy, which improves the transferability of supervised pre-training. Our work thus sheds light onto both the specific challenges of 3D transfer learning, as well as strategies to overcome them.

This paper introduces a pioneering approach to linearly controllable generative adversarial network (LC-GAN) driven by unsupervised learning. Departing from traditional methods relying on supervision signals or post-processing for latent feature disentanglement, our proposed technique enables unsupervised learning using only image data through contrastive feature categorization and spectral regularization. In our framework, the discriminator constructs geometry- and appearance-related feature spaces using a combination of image augmentation and contrastive representation learning. Leveraging these feature spaces, the generator autonomously categorizes input latent codes into geometry- and appearance-related features. Subsequently, the categorized features undergo projection into a subspace via our proposed spectral regularization, with each component adeptly controlling a distinct aspect of the generated image. Beyond providing fine-grained control over the generative model, our approach achieves state-of-the-art image generation quality on benchmark datasets, including FFHQ, CelebA-HQ, and AFHQ-V2.

To overcome the imbalanced multi-modal learning problem, where models prefer the training of specific modalities, existing methods propose to control the training of uni-modal encoders from different perspectives, taking the inter-modal performance discrepancy as the basis. However, the intrinsic limitation of modality capacity is ignored. The scarcely informative modalities are always recognized as worse-learnt'' ones in existing methods, which could force the model to memorize more noise, counterproductively affecting the multi-modal model ability. Moreover, the current modality modulation methods narrowly concentrate on selected worse-learnt modalities, even suppressing the training of others. Hence, it is essential to reasonably assess the learning state of each modality and take all modalities into account during balancing. To this end, we propose the Diagnosing & Re-learning method. The learning state of each modality is firstly estimated based on the separability of its uni-modal representation space, and then used to softly re-initialize the corresponding uni-modal encoder. In this way, encoders of worse learnt modalities are enhanced, simultaneously avoiding the over-training of other modalities. Accordingly, the multi-modal learning is effectively balanced and enhanced. Experiments covering multiple types of modalities and multi-modal frameworks demonstrate the superior performance of our simple-yet-effective method for balanced multi-modal learning.

Visual prompts represent a lightweight approach that adapts pre-trained models to downstream tasks without modifying the model parameters. They strategically transform the input and output through prompt engineering and label mapping, respectively. Yet, existing methodologies often overlook the synergy between these components, leaving the intricate relationship between them underexplored. To address this, we propose an Optimal Transport-based Label Mapping strategy (OTLM) that effectively reduces distribution migration and lessens the modifications required by the visual prompts. Specifically, we reconceptualize label mapping as a partial optimal transport problem, and introduce a novel transport cost matrix. Through the optimal transport framework, we establish a connection between output-side label mapping and input-side visual prompting. Additionally, we analyze frequency-based label mapping methods within this framework. We also offer an analysis of frequency-based label mapping techniques and demonstrate the superiority of our OTLM method. Our experiments across multiple datasets and various model architectures demonstrate significant performance improvements, which prove the effectiveness of the proposed method.

The pseudo-labelling algorithm is highly effective across various tasks, particularly in semi-supervised learning, yet its vulnerabilities are not always apparent on benchmark datasets, leading to suboptimal real-world performance. In this paper, we identified some channel activations in pseudo-labelling methods, termed disguising channel activations (abbreviated as disguising activations in the following sections), which exacerbate the confirmation bias issue when the training data distribution is inconsistent. Even state-of-the-art semi-supervised learning models exhibit significantly different levels of activation on some channels for data in different distributions, impeding the full potential of pseudo labelling. We take a novel perspective to address this issue by analysing the components of each channel's activation. Specifically, we model the activation of each channel as the mixture of two independent components. The mixture proportion enables us to identify the disguising activations, making it possible to employ our straightforward yet effective regularisation to attenuate the correlation between pseudo labels and disguising activations. This mitigation reduces the error risk of pseudo-label inference, leading to more robust optimization. The regularisation introduces no additional computing costs during the inference phase and can be seamlessly integrated as a plug-in into pseudo-labelling algorithms in various downstream tasks. Our experiments demonstrate that the proposed method achieves …

Large-scale image-text pre-trained models enable zero-shot classification and provide consistent accuracy across various data distributions. Nonetheless, optimizing these models in downstream tasks typically requires fine-tuning, which reduces generalization to out-of-distribution (OOD) data and demands extensive computational resources. We introduce Robust Adapter (R-Adapter), a novel method for fine-tuning zero-shot models to downstream tasks while simultaneously addressing both these issues. Our method integrates lightweight modules into the pre-trained model and employs novel self-ensemble techniques to boost OOD robustness and reduce storage expenses substantially. Furthermore, we propose MPM-NCE loss designed for fine-tuning on vision-language downstream tasks. It ensures precise alignment of multiple image-text pairs and discriminative feature learning. By extending the benchmark for robust fine-tuning beyond classification to include diverse tasks such as cross-modal retrieval and open vocabulary segmentation, we demonstrate the broad applicability of R-Adapter. Our extensive experiments demonstrate that R-Adapter achieves state-of-the-art performance across a diverse set of tasks, tuning only 13% of the parameters of the CLIP encoders.

Many leading self-supervised methods for unsupervised representation learning, in particular those for embedding image features, are built on variants of the instance discrimination task, whose optimization is known to be prone to instabilities that can lead to feature collapse. Different techniques have been devised to circumvent this issue, including the use of negative pairs with different contrastive losses, the use of external memory banks, and breaking of symmetry by using separate encoding networks with possibly different structures. Our method, termed BAM, rather than directly matching features of different views (augmentations) of input images, is based on matching their self-attention vectors, which are the distributions of similarities to the entire set of augmented images of a batch. We obtain rich representations and avoid feature collapse by minimizing a loss that matches these distributions to their globally balanced and entropy regularized version, which is obtained through a simple self-optimal-transport computation. We ablate and verify our method through a wide set of experiments that show competitive performance with leading methods on both semi-supervised and transfer-learning benchmarks. Our implementation and pre-trained models will be made publicly available.

Noisy few-shot learning (NFSL) presents novel challenges primarily due to the interplay between noisy labels and limited training data. While data cleansing offers a viable solution to address noisy labels in the general learning settings, it exacerbates information loss in FSL due to limited training data, resulting in inadequate model training. To best recover the underlying task manifold corrupted by the noisy labels, we resort to learning from uniquely designed unsupervised auxiliary tasks to compensate for information loss. Using unsupervised tasks can effectively avoid additional annotation costs and minimize the risk of introducing additional label noises. However, a randomly constructed unsupervised task may misguide the model to learn sample-specific features that are likely to compromise the primary few-shot learning task due to the noisy weak learning signals. We propose to conduct novel auxiliary task selection to ensure the intra-diversity among the unlabeled samples within a task. Domain invariant features are then learned from carefully constructed auxiliary tasks to best recover the original data manifold. We conduct theoretical analysis to derive novel generalization bounds for learning with auxiliary tasks. Extensive experiments are conducted to demonstrate that our method outperforms existing noisy few-shot learning methods under various in-domain and cross-domain few-shot classification …

One of the challenges for neural networks in real-life applications is the overconfident errors these models make when the data is not from the original training distribution. Addressing this issue is known as Out-of-Distribution (OOD) detection. Many state-of-the-art OOD methods employ an auxiliary dataset as a surrogate for OOD data during training to achieve improved performance. However, these methods fail to fully exploit the local information embedded in the auxiliary dataset. In this work, we propose the idea of leveraging the information embedded in the gradient of the loss function during training to enable the network to not only learn a desired OOD score for each sample but also to exhibit similar behavior in a local neighborhood around each sample data point. We also develop a novel energy-based sampling method to allow the network to be exposed to more informative OOD samples during the training phase. This is especially important when the auxiliary dataset is large. We demonstrate the effectiveness of our method through extensive experiments on several OOD benchmarks, improving the existing state-of-the-art FPR95 by 4% on our ImageNet experiment. We further provide a theoretical analysis through the lens of certified robustness and Lipschitz analysis to showcase the theoretical …

Recent studies highlight that deep learning models often learn spurious features mistakenly linked to labels, compromising their reliability in real-world scenarios where such correlations do not hold. Despite the increasing research effort, existing solutions often face two main challenges: they either demand substantial annotations of spurious attributes, or they yield less competitive outcomes with expensive training when additional annotations are absent. In this paper, we introduce SLIM, a cost-effective and performance-targeted approach to reducing spurious correlations in deep learning. Our method leverages a human-in-the-loop protocol featuring a novel attention labeling mechanism with a constructed attention representation space. SLIM significantly reduces the need for exhaustive additional labeling, requiring human input for fewer than 3% of instances. By prioritizing data quality over complicated training strategies, SLIM curates a smaller yet more feature-balanced data subset, fostering the development of spuriousness-robust models. Experimental validations across key benchmarks demonstrate that SLIM competes with or exceeds the performance of leading methods while significantly reducing costs. The SLIM framework thus presents a promising path for developing reliable models more efficiently.

Label dependencies have been widely studied in multi-label image recognition for improving performances. Previous methods mainly considered label co-occurrences as label correlations. In this paper, we show that label co-occurrences may be insufficient to represent label correlations, and modeling label correlations relies on latent context information. To this end, we propose a latent context embedding information network for multi-label image recognition. Our proposal is straightforward and contains three key modules to correspondingly tackle three questions, \ie, where to locate the latent context information, how to utilize the latent context information, and how to model label correlations with context-aware features. First, the multi-level context feature fusion module fuses the multi-level feature pyramids to obtain sufficient latent context information. Second, the latent context information embedding module aggregates the latent context information into categorical features, and thus the label correlation can be directly established. Moreover, we use the label correlation capturing module to model label correlations with full and partial manners, respectively. Comprehensive experiments validate the correctness of our arguments and the effectiveness of our method. In both generic multi-label classification and partial-label multi-label classification, our proposed method consistently achieves promising results.

Despite significant advancements in class-imbalanced semi-supervised learning (CISSL), many existing algorithms explicitly or implicitly assume that the class distribution of unlabeled data matches that of labeled data. However, when this assumption fails in practice, the classification performance of such algorithms may degrade due to incorrectly assigned weight to each class during training. We propose a novel CISSL algorithm called \textit{Rebalancing Using Estimated Class Distribution (RECD)}. RECD estimates the unknown class distribution of unlabeled data through Monte Carlo approximation, leveraging predicted class probabilities for unlabeled samples, and subsequently rebalances the classifier based on the estimated class distribution. Additionally, we propose an extension of feature clusters compression in the context of CISSL to mitigate feature map imbalance by densifying minority class clusters. Experimental results on four benchmark datasets demonstrate that RECD achieves state-of-the-art classification performance in CISSL.

Label noise is ubiquitous in real-world scenarios, posing a practical challenge to supervised models due to its effect in hurting the generalization performance of deep neural networks. Existing methods primarily employ the sample selection paradigm and usually rely on dataset-dependent prior knowledge (e.g., a pre-defined threshold) to cope with label noise, inevitably degrading the adaptivity. Moreover, existing methods tend to neglect the class balance in selecting samples, leading to biased model performance.To this end, we propose a simple yet effective approach named SED to deal with label noise in a Self-adaptivE and class-balanceD manner. Specifically, we first design a novel sample selection strategy to empower self-adaptivity and class balance when identifying clean and noisy data. A mean-teacher model is then employed to correct labels of noisy samples. Subsequently, we propose a self-adaptive and class-balanced sample re-weighting mechanism to assign different weights to detected noisy samples. Finally, we additionally employ consistency regularization on selected clean samples to improve model generalization performance. Extensive experimental results on synthetic and real-world datasets demonstrate the effectiveness and superiority of our proposed method.

We present a novel approach for generating minority samples that live on low-density regions of a data manifold. Our framework is built upon diffusion models, leveraging the principle of guided sampling that incorporates an arbitrary energy-based guidance during inference time. The key defining feature of our sampler lies in its self-contained nature, i.e., implementable solely with a pretrained model. This distinguishes our sampler from existing techniques that require expensive additional components (like external classifiers) for minority generation. Specifically, we first estimate the likelihood of features within an intermediate latent sample by evaluating a reconstruction loss w.r.t. its posterior mean. The generation then proceeds with the minimization of the estimated likelihood, thereby encouraging the emergence of minority features in the latent samples of subsequent timesteps. To further improve the performance of our sampler, we provide several time-scheduling techniques that properly manage the influence of guidance over inference steps. Experiments on benchmark real datasets demonstrate that our approach can greatly improve the capability of creating realistic low-likelihood minority instances over the existing techniques without the reliance on costly additional elements.

Novel Class Discovery (NCD) aims to discover unknown and novel classes in an unlabeled set by leveraging knowledge already learned about known classes. Existing works focus on instance-level or class-level knowledge representation and build a shared representation space to achieve performance improvements. However, a long-neglected issue is the potential imbalanced number of samples from known and novel classes, pushing the model towards dominant classes. Therefore, these methods suffer from a challenging trade-off between reviewing known classes and discovering novel classes. Based on this observation, we propose a Self-Cooperation Knowledge Distillation (SCKD) method to utilize each training sample (whether known or novel, labeled or unlabeled) for both review and discovery. Specifically, the model’s feature representations of known and novel classes are used to construct two disjoint representation spaces. Through spatial mutual information, we design a self-cooperation learning to encourage model learning from the two feature representation spaces from itself. Extensive experiments on six datasets demonstrate that our method can greatly enhance baseline performance, achieving competitive performance.

Existing approaches to Domain Incremental Learning (DIL) address catastrophic forgetting by storing and rehearsing exemplars from old domains. However, exemplar-based solutions are not always viable due to data privacy concerns or storage limitations. Therefore, Non-Exemplar Domain Incremental Learning (NEDIL) has emerged as a significant paradigm for resolving DIL challenges. Current NEDIL solutions extend the classifier incrementally for new domains to learn new knowledge, but unrestricted extension within the same feature space leads to inter-class confusion. To tackle this issue, we propose a simple yet effective method through cross-domain concePt INtegrAtion (PINA). We train a Unified Classifier (UC) as a concept container across all domains. Then, a Domain Specific Alignment (DSA) module is trained for each incremental domain, aligning the feature distribution to the base domain. During inference, we introduce a Patch Shuffle Selector (PSS) to select appropriate parameters of DSA for test images. Our developed patch shuffling technique disrupts class-dependent information, outperforming the domain selectors based on K-Nearest Neighbors or Nearest Mean Classifier. Extensive experiments demonstrate that our method achieves state-of-the-art performance while reducing the number of additional parameters. The source code will be released in http://XXX.XXX.XX.

Given a model trained on source data, Test-Time Adaptation (TTA) enables adaptation and inference in test data streams with domain shifts from the source. Current methods predominantly optimize the model for each incoming test data batch using self-training loss. While these methods yield commendable results in ideal test data streams, where batches are independently and identically sampled from the target distribution, they falter under more practical test data streams that are not independent and identically distributed (non-i.i.d.). The data batches in a non-i.i.d. stream display prominent label shifts relative to each other. It leads to conflicting optimization objectives among batches during the TTA process. Given the inherent risks of adapting the source model to unpredictable test-time distributions, we reverse the adaptation process and propose a novel Distribution Alignment loss for TTA. This loss guides the distributions of test-time features back towards the source distributions, which ensures compatibility with the well-trained source model and eliminates the pitfalls associated with conflicting optimization objectives. Moreover, we devise a domain shift detection mechanism to extend the success of our proposed TTA method in the continual domain shift scenarios. Our extensive experiments validate the logic and efficacy of our method. On six benchmark datasets, we …

Few-Shot Class-Incremental Learning (FSCIL) models aim to incrementally learn new classes with scarce samples while preserving knowledge of old ones. Existing FSCIL methods usually fine-tune the entire backbone, leading to overfitting and hindering the potential to learn new classes. On the other hand, recent prompt-based CIL approaches alleviate forgetting by training prompts with sufficient data in each task. In this work, we propose a novel framework named Attention-aware Self-adaptive Prompt (ASP). ASP encourages task-invariant prompts to capture shared knowledge by reducing specific information from the attention aspect. Additionally, self-adaptive task-specific prompts in ASP provide specific information and transfer knowledge from old classes to new classes with an Information Bottleneck learning objective. In summary, ASP prevents overfitting on base task and does not require enormous data in few-shot incremental tasks. Extensive experiments on three benchmark datasets validate that ASP consistently outperforms state-of-the-art FSCIL and prompt-based CIL methods in term of both learning new classes and mitigating forgetting.

Exemplar-free class-incremental learning using a backbone trained from scratch and starting from a small first task presents a significant challenge for continual representation learning. Prototype-based approaches, when continually updated, face the critical issue of semantic drift due to which the old class prototypes drift to different positions in the new feature space. Through an analysis of forgetting in prototype-based continual learning, we show that forgetting is not due to diminished discriminative power of the feature extractor, and can potentially be corrected by drift compensation. To address this, we propose a novel feature drift correction method called Learnable Drift Compensation (LDC). LDC can effectively mitigate drift in any moving backbone, whether supervised or unsupervised. LDC is fast and straightforward to integrate on top of existing continual learning approaches. Finally, we showcase how LDC can be applied in combination with self-supervised CL methods, resulting in the first exemplar-free semi-supervised continual learning approach. We achieve state-of-the-art performance in both supervised and semi-supervised settings across multiple datasets.

The open world is inherently dynamic, characterized by its ever-evolving concepts and distributions. Continual learning (CL) in this dynamic open-world environment, where knowledge must be continuously acquired from data streams without forgetting, presents a significant challenge. Existing CL methods, whether rehearsal-free or rehearsal-based, often struggle to effectively generalize to unseen test-time classes in this open-world context. To address this challenge, we introduce a new practical CL setting tailored for open-world visual representation learning. In this setting, subsequent data streams systematically introduce novel classes that are disjoint from the classes seen in previous training phases, all the while remaining distinct from the unseen test classes. In response, we introduce Dynamic Prompt and Representation Learner (DPaRL), a simple yet effective Prompt-based CL (PCL) method. Our DPaRL learns to generate dynamic prompts for inference, as opposed to relying on a static prompt pool in previous PCL methods. In addition, DPaRL jointly learns the dynamic prompt generation and discriminative representation at each training stage whereas prior PCL methods only refine the prompt learning throughout the process. Our experimental results demonstrate the superiority of our approach, surpassing state-of-the-art methods on well-established open-world image retrieval benchmarks by an average of 4.7% improvement in Recall@1 performance.

The rapid development of AI systems has been greatly influenced by the emergence of foundation models. A common approach for targeted problems involves fine-tuning these pre-trained foundation models for specific target tasks, resulting in a rapid spread of models fine-tuned across a diverse array of tasks. This work focuses on the problem of merging multiple fine-tunings of the same foundation model derived from a spectrum of auxiliary tasks. We introduce a new simple method, Model Breadcrumbs, which consists of a sparsely defined set of weights that carve out a trajectory within the weight space of a pre-trained model, enhancing task performance when traversed. These breadcrumbs are constructed by subtracting the weights from a pre-trained model before and after fine-tuning, followed by a sparsification process that eliminates weight outliers and negligible perturbations. Our experiments demonstrate the effectiveness of Model Breadcrumbs to simultaneously improve performance across multiple tasks. This contribution aligns with the evolving paradigm of updatable machine learning, reminiscent of the collaborative principles underlying open-source software development, fostering a community-driven effort to reliably update machine learning models. Our method is shown to be more efficient and unlike previous proposals does not require hyperparameter tuning for each new task added. Through extensive …

Knowledge distillation (KD) aims at improving the performance of a compact student model by distilling the knowledge from a high-performing teacher model. In this paper, we present an adaptive KD approach, namely AdaDistill, for deep face recognition. The proposed AdaDistill embeds the KD concept into softmax loss by training the student using a margin penalty softmax loss with distilled class centers from the teacher. Being aware of the relatively low capacity of the compact student model, we propose to distill relatively less complex knowledge at an early stage of training and more complex ones at a later stage of training. This relative adjustment of the distilled knowledge is controlled by the progression of the learning capability of the student over the training iterations without the need to tune any hyper-parameters. Extensive experiments and ablation studies prove that AdaDistill can enhance the discriminative learning capability of the student and demonstrate superiority over various state-of-the-art competitors on several challenging benchmarks such as IJB-B, IJB-C, and ICCV2021-MFR.

Deep learning benefits from the growing abundance of available data. Meanwhile, efficiently dealing with the growing data scale has become a challenge. Data publicly available are from different sources with various qualities, and it is impractical to do manual cleaning against noise and redundancy given today's data scale. There are existing techniques for cleaning/selecting the collected data. However, these methods are mainly proposed for offline settings that target one of the cleanness and redundancy problems. In practice, data are growing exponentially with both problems. This leads to repeated data curation with sub-optimal efficiency. To tackle this challenge, we propose InfoGrowth, an efficient online algorithm for data cleaning and selection, resulting in a growing dataset that keeps up to date with awareness of cleanliness and diversity. InfoGrowth can improve data quality/efficiency on both single-modal and multi-modal tasks, with an efficient and scalable design. Its framework makes it practical for real-world data engines.

Given a real-world dataset, data condensation (DC) aims to synthesize a small synthetic dataset that captures the knowledge of a natural dataset while being usable for training models with comparable accuracy. Recent works propose to enhance DC with data parameterization, which condenses data into very compact parameterized data containers instead of images. By optimizing with an appropriate loss function, data parameterization methods can generate high-quality synthetic datasets and achieve improved model performance. Top-performing data parameterization methods use GPU memory intensive trajectory-based losses in their optimization. In this paper, we propose a novel data parameterization architecture, Hierarchical Memory Network (HMN), that achieves comparable or better performance to SOTA methods, even with a GPU memory friendly batch-based loss function. HMN's key insight is to directly capture sharing of features at both within-class level and across-class level by proposing a hierarchical parameterized architecture. We evaluate HMN on five public datasets and show that HMN outperforms current baselines (including those using trajectory-based losses), even when HMNs are trained with a GPU-friendly batch-based loss.

Deploying models across diverse devices demands tradeoffs among multiple objectives due to different resource constraints. Arguably, due to the small model trap problem in multi-objective neural architecture search (MO-NAS) based on a supernet, existing approaches may fail to maintain large models. Moreover, multi-tasking neural architecture search (MT-NAS) excels in handling multiple tasks simultaneously, but most existing efforts focus on tasks from the same dataset, limiting their practicality in real-world scenarios where multiple tasks may come from distinct datasets. To tackle the above challenges, we propose a Multi-Objective Evolutionary Multi-Tasking framework for NAS (MO-EMT-NAS) to achieve architectural knowledge transfer across tasks from different datasets while finding Pareto optimal architectures for multi-objectives, model accuracy and computational efficiency. To alleviate the small model trap issue, we introduce an auxiliary objective that helps maintain multiple larger models of similar accuracy. Moreover, the computational efficiency is further enhanced by parallelizing the training and validation of the weight-sharing-based supernet. Experimental results on seven datasets with two, three, and four task combinations show that MO-EMT-NAS achieves a better minimum classification error while being able to offer flexible trade-offs between model performance and complexity, compared to the state-of-the-art single-objective MT-NAS algorithms. In addition, the runtime of MO-EMT-NAS is …

Federated learning (FL) is a general principle for decentralized clients to train a server model collectively without sharing local data. FL is a promising framework with practical applications, but its standard training paradigm requires the clients to backpropagate through the model to compute gradients. Since these clients are typically edge devices and not fully trusted, executing backpropagation on them incurs computational and storage overhead as well as white-box vulnerability. In light of this, we develop backpropagation-free federated learning, dubbed BAFFLE, in which backpropagation is replaced by multiple forward processes to estimate gradients. BAFFLE is 1) memory-efficient and easily fits uploading bandwidth; 2) compatible with inference-only hardware optimization and model quantization or pruning; and 3) well-suited to trusted execution environments, because the clients in BAFFLE only execute forward propagation and return a set of scalars to the server. Empirically we use BAFFLE to train deep models from scratch or to finetune pretrained models, achieving acceptable results.

Attribution-based explanations are garnering increasing attention recently and have emerged as the predominant approach towards \textit{eXplanable Artificial Intelligence}~(XAI). However, the absence of consistent configurations and systematic investigations in prior literature impedes comprehensive evaluations of existing methodologies. In this work, we introduce {Meta-Rank}, an open platform for benchmarking attribution methods in the image domain. Presently, Meta-Rank assesses eight exemplary attribution methods using six renowned model architectures on four diverse datasets, employing both the \textit{Most Relevant First} (MoRF) and \textit{Least Relevant First} (LeRF) evaluation protocols.Through extensive experimentation, our benchmark reveals three insights in attribution evaluation endeavors: 1) evaluating attribution methods under disparate settings can yield divergent performance rankings; 2) although inconsistent across numerous cases, the performance rankings exhibit remarkable consistency across distinct checkpoints along the same training trajectory; 3) prior attempts at consistent evaluation fare no better than baselines when extended to more heterogeneous models and datasets. Our findings underscore the necessity for future research in this domain to conduct rigorous evaluations encompassing a broader range of models and datasets, and to reassess the assumptions underlying the empirical success of different attribution methods. Code, models, and datasets will be made publicly available in the near future.

Nowadays an ever-growing concerning phenomenon, the emergence of algorithmic biases that can lead to unfair models, emerges. Several debiasing approaches have been proposed in the realm of deep learning, employing more or less sophisticated approaches to discourage these models from massively employing these biases. However, a question emerges: is this extra complexity really necessary? Is a vanilla-trained model already embodying some unbiased sub-networks'' that can be used in isolation and propose a solution without relying on the algorithmic biases? In this work, we show that such a sub-network typically exists, and can be extracted from a vanilla-trained model without requiring additional training. We further validate that such specific architecture is incapable of learning a specific bias, suggesting that there are possible architectural countermeasures to the problem of biases in deep neural networks.

Neural networks are susceptible to adversarial perturbations that are transferable across different models. In this paper, we introduce a novel model alignment technique aimed at improving a given source model's ability in generating transferable adversarial perturbations. During the alignment process, the parameters of the source model are fine-tuned to minimize an alignment loss. This loss measures the divergence in the predictions between the source model and another, independently trained model, referred to as the witness model. To understand the effect of model alignment, we conduct a geometric analysis of the resulting changes in the loss landscape. Extensive experiments on the ImageNet dataset, using a variety of model architectures, demonstrate that perturbations generated from aligned source models exhibit significantly higher transferability than those from the original source model.

While the success of deep learning relies on large amounts of training datasets, data is often limited in privacy-sensitive domains. To address this challenge, generative model learning with differential privacy has emerged as a solution to train private generative models for desensitized data generation. However, the quality of the images generated by existing methods is limited due to the complexity of modeling data distribution. We build on the success of diffusion models and introduce DP-SAD, which trains a private diffusion model by a stochastic adversarial distillation method. Specifically, we first train a diffusion model as a teacher and then train a student by distillation, in which we achieve differential privacy by adding noise to the gradients from other models to the student. For better generation quality, we introduce a discriminator to distinguish whether an image is from the teacher or the student, which forms the adversarial training. Extensive experiments and analysis clearly demonstrate the effectiveness of our proposed method.

Recent model inversion attack algorithms permit adversaries to reconstruct a neural network's private training data just by repeatedly querying the network and inspecting its outputs. In this work, we develop a novel network architecture that leverages sparse-coding layers to obtain superior robustness to this class of attacks. Three decades of computer science research has studied sparse coding in the context of image denoising, object recognition, and adversarial misclassification settings, but to the best of our knowledge, its connection to state-of-the-art privacy vulnerabilities remains unstudied. However, sparse coding architectures suggest an advantageous means to defend against model inversion attacks because they allow us to control the amount of irrelevant private information encoded in a network's intermediate representations in a manner that can be computed efficiently during training and that is known to have little effect on classification accuracy. Specifically, compared to networks trained with a variety of state-of-the-art defenses, our sparse-coding architectures maintain comparable or higher classification accuracy while degrading state-of-the-art training data reconstructions by factors of 1.1 to 18.3 across a variety of reconstruction quality metrics (PSNR, SSIM, FID). This performance advantage holds across 5 datasets ranging from CelebA faces to medical images and MNIST, and across various state-of-the-art SGD-based …

Diffusion Models (DMs) achieve state-of-the-art synthesis results in image generation and have been applied to various fields. However, DMs sometimes seriously violate user privacy during usage, making the protection of privacy an urgent issue. Using traditional privacy computing schemes like Secure Multi-Party Computation (MPC) directly in DMs faces significant computation and communication challenges. To address these issues, we propose CipherDM, the first novel, versatile and universal framework applying MPC technology to DMs for secure sampling, which can be widely implemented on multiple DM based tasks. We thoroughly analyze sampling latency breakdown, find the time-consuming part and design corresponding secure MPC protocols for computing nonlinear activations including SoftMax, SiLU and Mish. CipherDM is evaluated on popular architectures (DDPM, DDIM) using the MNIST dataset and on SD deployed by diffusers. Compared to direct implementation on SPU, our approach improves running time by approximately 1.084× ∼ 2.328×, and reduces communication costs by approximately 1.212× ∼ 1.791×. Code will be available upon paper acceptance.

Deep neural networks (DNNs) have demonstrated effectiveness in various fields. However, DNNs are vulnerable to backdoor attacks, which inject a unique pattern, called trigger, in the input to cause misclassification to an attack-chosen target label. While existing works have proposed various methods to mitigate backdoor effects in poisoned models, they tend to be less effective against recent advanced attacks. In this paper, we introduce a novel post-training defense technique UNIT that can effectively remove backdoors for a variety of attacks. In specific, UNIT approximates a unique and tight activation distribution for each neuron in the model. It then proactively dispels substantially large activation values that exceed the approximated boundaries. Our experimental results demonstrate that UNIT outperforms 9 popular defense methods against 14 existing backdoor attacks, including 2 advanced attacks, using only 5\% of clean training data.

We propose StereoGlue, a method designed for joint feature matching and robust estimation that effectively reduces the combinatorial complexity of these tasks using single-point minimal solvers. StereoGlue is applicable to a range of problems, including but not limited to relative pose and homography estimation, determining absolute pose with 2D-3D correspondences, and estimating 3D rigid transformations between point clouds. StereoGlue starts with a set of one-to-many tentative correspondences, iteratively forms tentative matches, and estimates the minimal sample model. This model then facilitates guided matching, leading to consistent one-to-one matches, whose number serves as the model score. StereoGlue is superior to the state-of-the-art robust estimators on real-world datasets on multiple problems, improving upon a number of recent feature detectors and matchers. Additionally, it shows improvements in point cloud matching and absolute camera pose estimation. The code will be made publicly available.

Recent advances in point cloud registration mostly leverage geometric information. Although these methods have yielded promising results, they still struggle with problems of low overlap, thus limiting their practical usage. In this paper, we propose ML-SemReg, a plug-and-play point cloud registration framework that fully exploits semantic information. Our key insight is that mismatches can be categorized into two types, i.e., inter- and intra-class, after rendering semantic clues, and can be well addressed by utilizing multi-level semantic consistency. We first propose a Semantic Cluster Matching module to address inter-class mismatching, outputting multiple matching groups that inherently satisfy Neighborhood Semantic Consistency. For each group, a Semantic Mask Matching module based on Scene Semantic Consistency is then introduced to suppress intra-class mismatching. Benefit from those two modules, ML-SemReg generates correspondences with a high inlier ratio. Extensive experiments demonstrate excellent performance and robustness of ML-SemReg, e.g., in hard-cases of the KITTI dataset, the Registration Recall of MAC increases by almost 34 percentage points when our ML-SemReg is equipped. Code is available at \url{https://github.com/qwAyu/ML-SemReg}

Humans naturally build mental models of object interactions and dynamics, allowing them to imagine how their surroundings will change if they take a certain action. While generative models today have shown impressive results on generating/editing images unconditionally or conditioned on text, current methods do not provide the ability to perform fine-grained object manipulation conditioned on actions, an important tool for world modeling and action planning. Therefore, we propose learning to model interactions through a novel form of visual conditioning: hands. Hands are a natural way to specify control through actions such as grasping, pulling, pushing, etc. Given an input image and a representation of a hand interacting with the scene, our approach, CoSHAND, synthesizes a depiction of what the scene would look like after the interaction has occurred. We show that CoSHAND is able to recover the dynamics of manipulation by learning from large amounts of unlabeled videos of human hands interacting with objects, and leveraging internet-scale latent diffusion model priors. The model demonstrates strong capabilities on a variety of actions and object types beyond the dataset, and the ability to generate multiple possible futures depending on the actions performed. CoSHAND is also able to generalize zero-shot to tasks where …

Diffusion models have shown remarkable results in generating 2D images and small-scale 3D objects. However, their application to the synthesis of large-scale 3D scenes has been rarely explored. This is mainly due to the inherent complexity and bulky size of 3D scenery data, particularly outdoor scenes, and the limited availability of comprehensive real-world datasets, which makes training a stable scene diffusion model challenging. In this work, we explore how to effectively generate large-scale 3D scenes using the coarse-to-fine paradigm. We introduce a framework, the Pyramid Discrete Diffusion model (PDD), which employs scale-varied diffusion models to progressively generate high-quality outdoor scenes. Experimental results of PDD demonstrate our successful exploration in generating 3D scenes both unconditionally and conditionally. We further showcase the data compatibility of the PDD model, due to its multi-scale architecture: a PDD model trained on one dataset can be easily fine-tuned with another dataset. The source codes and trained models will be made available to the public.

While personalized text-to-image generation has enabled the learning of a single concept from multiple images, a more practical yet challenging scenario involves learning multiple concepts within a single image. However, existing works tackling this scenario heavily rely on extensive human annotations. In this paper, we introduce a novel task named Unsupervised Concept Extraction (UCE) that considers an unsupervised setting without any human knowledge of the concepts. Given an image that contains multiple concepts, the task aims to extract and recreate individual concepts solely relying on the existing knowledge from pretrained diffusion models. To achieve this, we present ConceptExpress that tackles UCE by unleashing the inherent capabilities of pretrained diffusion models in two aspects. Specifically, a concept localization approach automatically locates and disentangles salient concepts by leveraging spatial correspondence from diffusion self-attention; and based on the lookup association between a concept and a conceptual token, a concept-wise optimization process learns discriminative tokens that represent each individual concept. Finally, we establish an evaluation protocol tailored for the UCE task. Extensive experiments demonstrate that ConceptExpress is a promising solution to the UCE task.

Recently, various methods have been proposed to address the inconsistency issue of DDIM inversion to enable image editing, such as EDICT [39] and Null-text inversion [23]. However, the above methods introduce considerable computational overhead. In this paper, we propose a new technique, named bi-directional integration approximation (BDIA), to perform exact diffusion inversion with negligible computational overhead. Suppose we would like to estimate the next diffusion state z{i-1} at timestep ti with the historical information (i, zi) and (i+1, z{i+1}). We first obtain the estimated Gaussian noise epsilon(zi, i), and then apply the DDIM update procedure twice for approximating the ODE integration over the next time-slot [ti, t{i-1}] in the forward manner and the previous time-slot [ti, t{t+1}] in the backward manner. The DDIM step for the previous time-slot is used to refine the integration approximation made earlier when computing zi. A nice property of BDIA-DDIM is that the update expression for z{i-1} is a linear combination of (z{i+1}, zi, epsilon(zi, i)). This allows for exact backward computation of z{i+1} given (zi, z_{i-1}), thus leading to exact diffusion inversion. We perform a convergence analysis for …

Recent developments in diffusion models have advanced conditioned image generation, yet they struggle with reconstructing out-of-distribution (OOD) images, such as unseen tumors, causing 'image hallucination' and risking misdiagnosis. We hypothesize such hallucinations result from local OOD regions in the conditional images. By partitioning the OOD region and conducting separate generations, hallucinations can be alleviated, and we verify this with motivational studies in several applications. From this, we propose a training-free diffusion framework that reduces hallucination by performing multiple \textit{Local Diffusion} processes. Our approach involves OOD estimation followed by two diffusion modules: a 'branching' module for local image generations from OOD estimations, and a 'fusion' module to integrate these predictions into a full image cohesively. These modules adapt to each testing dataset by updating an auxiliary classifier. Our evaluation shows our method improves baseline models quantitatively and qualitatively across different datasets. It also works well with various pre-trained diffusion models as a plug-and-play option.

Recent studies on inverse problems have proposed posterior samplers that leverage the pre-trained diffusion models as a powerful prior. The attempts have paved the way for using diffusion models in a wide range of inverse problems. However, the existing methods entail computationally demanding iterative sampling procedures and optimize a separate solution for each measurement, which leads to limited scalability and lack of generalization capability across unseen samples. To address these limitations, we propose a novel approach, Diffusion prior-based Amortized Variational Inference (DAVI) that solves inverse problems with a diffusion prior from an amortized variational inference perspective. Specifically, instead of the separate measurement-wise optimization, our amortized inference learns a function that directly maps measurements to the implicit posterior distributions of corresponding clean data, enabling a single-step posterior sampling even for unseen measurements. The proposed method learns the function by minimizing the Kullback-Leibler divergence between the implicit distributions and the true posterior distributions with multiple measurements using objectives derived based on variational inference. Extensive experiments across three image restoration tasks, e.g., Gaussian deblur, 4x super-resolution, and box inpainting with two benchmark datasets, demonstrate our superior performance over strong diffusion model-based methods.

We introduce Adversarial Diffusion Distillation (ADD), a novel training approach that efficiently samples large-scale foundational image diffusion models in just 1--4 steps while maintaining high image quality. We use score distillation to leverage large-scale off-the-shelf image diffusion models as a teacher signal in combination with an adversarial loss to ensure high image fidelity even in the low-step regime of one or two sampling steps. Our analyses show that our model clearly outperforms existing few-step methods (GANs, Latent Consistency Models) in a single step and reaches the performance of state-of-the-art diffusion models (SDXL) in only four steps. ADD is the first method to unlock single-step, real-time image synthesis with foundation models.

This paper presents Arc2Face, an identity-conditioned face foundation model, which, given the ArcFace embedding of a person, can generate diverse photo-realistic images with an unparalleled degree of face similarity than existing models. Despite previous attempts to decode face recognition features into detailed images, we find that common high-resolution datasets (e.g. FFHQ) lack sufficient identities to reconstruct any subject. To that end, we meticulously upsample a significant portion of the WebFace42M database, the largest public dataset for face recognition (FR). Arc2Face builds upon a pretrained Stable Diffusion model, yet adapts it to the task of ID-to-face generation, conditioned solely on ID vectors. Deviating from recent works that combine ID with text embeddings for zero-shot personalization of text-to-image models, we emphasize on the compactness of FR features, which can fully capture the essence of the human face, as opposed to hand-crafted prompts. Crucially, text-augmented models struggle to decouple identity and text, usually necessitating some description of the given face to achieve satisfactory similarity. Arc2Face, however, only needs the discriminative features of ArcFace to guide the generation, offering a robust prior for a plethora of tasks where ID consistency is of paramount importance. As an example, we train a FR model on synthetic …

Federated Class Continual Learning (FCCL) merges the challenges of distributed client learning with the need for seamless adaptation to new classes without forgetting old ones. The key challenge in FCCL is catastrophic forgetting, an issue that has been explored to some extent in Continual Learning (CL). However, due to privacy preservation requirements, some conventional methods, such as experience replay, are not directly applicable to FCCL. Existing FCCL methods mitigate forgetting by generating historical data through federated training of GANs or data-free knowledge distillation. However, these approaches often suffer from unstable training of generators or low-quality generated data, limiting their guidance for the model. To address this challenge, we propose a novel method of data replay based on diffusion models. Instead of training a diffusion model, we employ a pre-trained conditional diffusion model to reverse-engineer each category, searching the corresponding input conditions for each category within the model's input space, significantly reducing computational resources and time consumption while ensuring effective generation. Furthermore, we enhance the classifier's domain generalization ability on generated and real data through contrastive learning, indirectly improving the representational capability of generated data for real data. Extensive experiments demonstrate that our method significantly outperforms existing baselines.

Omnidirectional images (ODIs) are commonly used in real-world visual tasks, and high-resolution ODIs help improve the performance of related visual tasks. Most existing super-resolution methods for ODIs use end-to-end learning strategies, resulting in inferior realness of generated images and a lack of effective out-of-domain generalization capabilities in training methods. Image generation methods represented by diffusion model provide strong priors for visual tasks and have been proven to be effectively applied to image restoration tasks. Leveraging the image priors of the Stable Diffusion (SD) model, we achieve omnidirectional image super-resolution with both fidelity and realness, dubbed as OmniSSR. Firstly, we transform the equirectangular projection (ERP) images into tangent projection (TP) images, whose distribution approximates the planar image domain. Then, we use SD to iteratively sample initial high-resolution results. At each denoising iteration, we further correct and update the initial results using the proposed Octadecaplex Tangent Information Interaction (OTII) and Gradient Decomposition (GD) technique to ensure better consistency. Finally, the TP images are transformed back to obtain the final high-resolution results. Our method is zero-shot, requiring no training or fine-tuning. Experiments of our method on two benchmark datasets demonstrate the effectiveness of our proposed method.

Spatio-temporal video grounding aims to localize the spatio-temporal tube in a video according to the given language query. To eliminate the annotation costs, we make a first exploration to tackle spatio-temporal video grounding in a zero-shot manner. Our method dispenses with the need for any training videos or annotations; instead, it localizes the target object by leveraging large visual-language models and optimizing within the video and text query during the test time. To enable spatio-temporal comprehension, we introduce a multimodal modulation that integrates the spatio-temporal context into both visual and textual representation. On the visual side, we devise a context-based visual modulation that amplifies the visual representation by propagation and aggregation of the contextual semantics. Concurrently, on the textual front, we propose a prototype-based textual modulation to refine the textual representations using visual prototypes, effectively mitigating the cross-modal discrepancy. In addition, to overcome the interleaved spatio-temporal dilemma, we propose an expectation maximization (EM) framework to optimize the process of temporal relevance estimation and spatial region identification in an alternating way. Comprehensive experiments validate that our zero-shot approach achieves superior performance in comparison to several state-of-the-art methods with stronger supervision. We will make our code publicly accessible online.

We present a method to build animatable dog avatars from monocular videos. This is challenging as animals display a range of (unpredictable) non-rigid movements and have a variety of appearance details (e.g., fur, spots, tails). We develop an approach that links the video frames via a 4D solution that jointly solves for animal's pose variation, and its appearance (in a canonical pose). To this end, we significantly improve the quality of template-based shape fitting by endowing the SMAL parametric model with Continuous Surface Embeddings (CSE), which brings image-to-mesh reprojection constaints that are denser, and thus stronger, than the previously used sparse semantic keypoint correspondences. To model appearance, we propose an implicit duplex-mesh texture that is defined in the canonical pose, but can be deformed using SMAL pose coefficients and later rendered to enforce a photometric compatibility with the input video frames. On the challenging CoP3D and APTv2 datasets, we demonstrate superior results (both in terms of pose estimates and predicted appearance) to existing template-free (RAC) and template-based approaches (BARC, BITE)

Our objective is to discover and localize monotonic temporal changes in a sequence of images. To achieve this, we exploit a simple proxy task of ordering a shuffled image sequence, with `time' serving as a supervisory signal since only changes that are monotonic with time can give rise to the correct ordering. We also introduce a flexible transformer-based model for general-purpose ordering of image sequences of arbitrary length with built-in attribution maps. After training, the model successfully discovers and localizes monotonic changes while ignoring cyclic and stochastic ones. We demonstrate applications of the model in multiple video settings covering different scene and object types, discovering both object-level and environmental changes in unseen sequences. We also demonstrate that the attention-based attribution maps function as effective prompts for segmenting the changing regions, and that the learned representations can be used for downstream applications. Finally, we show that the model achieves the state of the art on standard benchmarks for ordering a set of images.

Action Quality Assessment (AQA) evaluates diverse skills but models struggle with non-stationary data. We propose Continual AQA (CAQA) to refine models using sparse new data. Feature replay preserves memory without storing raw inputs. However, the misalignment between static old features and the dynamically changing feature manifold causes severe catastrophic forgetting. To address this novel problem, we propose Manifold-Aligned Graph Regularization (MAGR), which first aligns deviated old features to the current feature manifold, ensuring representation consistency. It then constructs a graph jointly arranging old and new features aligned with quality scores. Experiments show MAGR outperforms recent strong baselines with up to 6.56%, 5.66%, 15.64%, and 9.05% correlation gains on the MTL-AQA, FineDiving, UNLV-Dive, and JDM-MSA split datasets, respectively. This validates MAGR for continual assessment challenges arising from non-stationary skill variations.

Compositional actions consist of dynamic (verbs) and static (objects) concepts. Humans can easily recognize unseen compositions using the learned concepts. For machines, solving such a problem requires a model to recognize unseen actions composed of previously observed verbs and objects, thus requiring, so-called, compositional generalization ability. To facilitate this research, we propose a novel Zero-Shot Compositional Action Recognition (ZS-CAR) task. For evaluating the task, we construct a new benchmark, Something-composition (Sth-com), based on the widely used Something-Something V2 dataset. We also propose a novel Component-to-Composition (C2C) learning method to solve the new ZS-CAR task. C2C includes an independent component learning module and a composition inference module. Last, we devise an enhanced training strategy to address the challenges of component variation between seen and unseen compositions and to handle the subtle balance between learning seen and unseen actions. The experimental results demonstrate that the proposed framework significantly surpasses the existing compositional generalization methods and sets a new state-of-the-art. The new Sth-com benchmark and code are available at https://anonymous.4open.science/r/C2C_anonymous-51F1.

Empowered by Large Language Models (LLMs), recent advancements in VideoLLMs have driven progress in various video understanding tasks. These models encode video representations through pooling or query aggregation over a vast amount of visual tokens, making computational and memory costs affordable. Despite successfully providing an overall comprehension of video content, existing VideoLLMs still face challenges in achieving detailed understanding in videos due to overlooking local information in long-term videos. To tackle this challenge, we introduce LongVLM, a straightforward yet powerful VideoLLM for long video understanding, building upon the observation that long videos often consist of sequential key events, complex actions, and camera movements. Our approach proposes to decompose long videos into multiple short-term segments and encode local features for each local segment via a hierarchical token merging module. These features are concatenated in temporal order to maintain the storyline across sequential short-term segments. Additionally, we propose to integrate global semantics into each local feature to enhance context understanding. In this way, we encode video representations that incorporate both local and global information, enabling the LLM to generate comprehensive responses for long-term videos. Experimental results on the VideoChatGPT benchmark and zero-shot video question-answering datasets demonstrate the superior capabilities of our model …

Goal-oriented planning, or anticipating a series of actions that transition an agent from its current state to a predefined objective, is crucial for developing intelligent assistants aiding users in daily procedural tasks. The problem presents significant challenges due to the need for comprehensive knowledge of temporal and hierarchical task structures, as well as strong capabilities in reasoning and planning. To achieve this, prior work typically relies on extensive training on the target dataset, which often results in significant dataset bias and a lack of generalization to unseen tasks. In this work, we introduce VidAssist, an integrated framework designed for zero/few-shot goal-oriented planning in instructional videos. VidAssist leverages large language models (LLMs) as both the knowledge base and the assessment tool for generating and evaluating action plans, thus overcoming the challenges of acquiring procedural knowledge from small-scale, low-diversity datasets. Moreover, VidAssist employs a breadth-first search algorithm for optimal plan generation, in which a composite of value functions designed for goal-oriented planning are utilized to assess the predicted actions at each step. Extensive experiments demonstrate that VidAssist offers a unified framework for different goal-oriented planning setups, e.g., visual planning for assistance (VPA) and procedural planning (PP) and achieves remarkable performance in zero-shot …

The unprecedented surge in video data production in recent years necessitates efficient tools for extracting meaningful frames from videos for downstream tasks. Long-term temporal reasoning is a key desideratum for frame retrieval systems. While state-of-the-art foundation models, like VideoLLaMA and ViCLIP, are proficient in short-term semantic understanding, they surprisingly fail at long-term reasoning across frames. A key reason for their failure is that they intertwine per-frame perception and temporal reasoning into a single deep network. Hence, decoupling but co-designing semantic understanding and temporal reasoning is essential for efficient scene identification. We propose a system that leverages vision-language models for semantic understanding of individual frames but effectively reasons about the long-term evolution of events using state machines and temporal logic (TL) formulae that inherently capture memory. Our TL-based reasoning improves the F1 score of complex event identification by 9-15% compared to benchmarks that use GPT4 for reasoning on state-of-the-art self-driving datasets such as Waymo and NuScenes.

Video action detection (VAD) aims to detect actors and classify their actions in a video. We figure that VAD suffers more from classification rather than localization of actors. Hence, we analyze how prevailing methods form features for classification and find that they prioritize actor regions for classification, yet often overlooking the essential contextual information necessary for accurate classification. Accordingly, we propose to reduce the model's bias toward the actor itself and encourage it to pay attention to the context that is relevant to each action class. By assigning a class-dedicated query to each action class, the model can dynamically determine where to focus for effective classification. The proposed method demonstrates superior performance on three challenging benchmarks while using significantly fewer parameters and less computation.

Video recognition models often learn scene-biased action representation due to the spurious correlation between actions and scenes in the training data. Such models show poor performance when the test data consists of videos with unseen action-scene combinations. Although Scene-debiased action recognition models might address the issue, they often overlook valuable scene information in the data. To address this challenge, we propose to learn Disentangled VIdeo representations of Action and Scene (DEVIAS), for more holistic video understanding. We propose an encoder-decoder architecture to learn disentangled action and scene representations with a single model. The architecture consists of a disentangling encoder (DE), an action mask decoder (AMD), and a prediction head. The key to achieving the disentanglement is employing both DE and AMD during training time. The DE uses the slot attention mechanism to learn disentangled action and scene representations. For further disentanglement, an AMD learns to predict action masks, given an action slot. With the resulting disentangled representations, we can achieve robust performance across diverse scenarios, including both seen and unseen action-scene combinations. We rigorously validate the proposed method on the UCF-101, Kinetics-400, and HVU datasets for the seen, and the SCUBA, HAT, and HVU datasets for unseen action-scene combination scenarios. …

Temporal video alignment aims to synchronize the key events like object interactions or action phase transitions in two videos. Such methods could benefit various video editing, processing, and understanding tasks. However, existing approaches operate under the restrictive assumption that a suitable video pair for alignment is given, significantly limiting their broader applicability. To address this, we re-pose temporal alignment as a search problem and introduce the task of Alignable Video Retrieval (AVR). Given a query video, our approach can identify well-alignable videos from a large collection of clips and temporally synchronize them to the query. To achieve this, we make three key contributions: 1) we introduce DRAQ, a video alignability indicator to identify and re-rank the best alignable video from a set of candidates; 2) we propose an effective and generalizable frame-level video feature design to improve the alignment performance of several off-the-shelf feature representations, and 3) we propose a novel benchmark and evaluation protocol for AVR using cycle-consistency metrics. Our experiments on 3 datasets, including large-scale Kinetics700, demonstrate the effectiveness of our approach in identifying alignable video pairs from diverse datasets.

This paper proposes a simple, yet effective framework, called GiT, simultaneously applicable for various vision tasks only with a vanilla ViT. Motivated by the universality of the Multi-layer Transformer architecture (e.g., GPT) widely used in large language models (LLMs), we seek to broaden its scope to serve as a powerful vision foundation model (VFM). However, unlike language modeling, visual tasks typically require specific modules, such as bounding box heads for detection and pixel decoders for segmentation, greatly hindering the application of powerful multi-layer transformers in the vision domain. To solve this, we design a universal language interface that empowers the successful auto-regressive decoding to adeptly unify various visual tasks, from image-level understanding (e.g., captioning), over sparse perception (e.g., detection), to dense prediction (e.g., segmentation). Based on the above designs, the entire model is composed solely of a ViT, without any specific additions, offering a remarkable architectural simplification. GiT is a multi-task visual model, jointly trained across five representative benchmarks without task-specific fine-tuning. Interestingly, our GiT builds a new benchmark in generalist performance, and fosters mutual enhancement across tasks, leading to significant improvements compared to isolated training. This reflects a similar impact observed in LLMs. Further enriching training with 27 datasets, …

Vision-Language Pre-training (VLP) models like CLIP have achieved remarkable success in computer vision and particularly demonstrated superior robustness to distribution shifts of 2D images. However, their robustness under 3D viewpoint variations is still limited, which can hinder the development for real-world applications. This paper successfully addresses this concern while keeping VLPs' original performance by breaking through two primary obstacles: 1) the scarcity of training data and 2) the suboptimal fine-tuning paradigms. To combat data scarcity, we build the Multi-View Caption (MVCap) dataset --- a comprehensive collection of over four million multi-view image-text pairs across more than 100K objects, providing more potential for VLP models to develop generalizable viewpoint-invariant representations. To address the limitations of existing paradigms in performance trade-offs and training efficiency, we design a novel fine-tuning framework named Omniview-Tuning (OVT). Specifically, OVT introduces a Cross-Viewpoint Alignment objective through a minimax-like optimization strategy, which effectively aligns representations of identical objects from diverse viewpoints without causing overfitting. Additionally, OVT fine-tunes VLP models in a parameter-efficient manner, leading to minimal computational cost. Extensive experiments on various VLP models with different architectures validate that OVT significantly improves the models' resilience to viewpoint shifts and keeps the original performance, establishing a pioneering standard for …

Vision-Language Large Models (VLMs) recently become primary backbone of AI, due to the impressive performance. However, their expensive computation costs, i.e., throughput and delay, impede potentials in the real-world scenarios. To achieve acceleration for VLMs, most existing methods focus on the model perspective: pruning, distillation, quantization, but completely overlook the data-perspective redundancy. To fill the overlook, this paper pioneers the severity of data redundancy, and designs one plug-and-play Turbo module guided by information degree to prune inefficient tokens from visual or textual data. In pursuit of efficiency-performance trade-offs, information degree takes two crucial factors into consideration: mutual redundancy and semantic value. Concretely, the former evaluates data duplication between sequential tokens; while the latter evaluates each token by its contribution to the overall semantics. As a result, tokens with high information degree carry less redundancy and stronger semantics. For VLMs' calculation, Turbo works as a user-friendly plug-in that sorts data referring to information degree, utilizing only top-level ones to save costs. Its advantages are multifaceted, e.g., being generally compatible to various VLMs across understanding and generation, simple use without retraining and trivial engineering efforts. On multiple VLMs benchmarks, we fully experiment to reveal good acceleration of Turbo, under negligible performance drop.

Large vision-language models (VLMs) have recently achieved remarkable progress, exhibiting impressive multimodal perception and reasoning abilities. However, effectively evaluating these large VLMs remains a major challenge, hindering future development in this domain. Traditional benchmarks like VQAv2 or COCO Caption provide quantitative performance measurements but lack fine-grained ability assessment and robust evaluation metrics. Meanwhile, subjective benchmarks, such as OwlEval, offer comprehensive evaluations of a model's abilities by incorporating human labor, which is not scalable and may display significant bias. In response to these challenges, we propose MMBench, a bilingual benchmark for assessing the multi-modal capabilities of VLMs. MMBench methodically develops a comprehensive evaluation pipeline, primarily comprised of the following key features: 1. MMBench is meticulously curated with well-designed quality control schemes, surpassing existing similar benchmarks in terms of the number and variety of evaluation questions and abilities; 2. MMBench introduces a rigorous CircularEval strategy and incorporates large language models to convert free-form predictions into pre-defined choices, which helps to yield accurate evaluation results for models with limited instruction-following capabilities. 3. MMBench incorporates multiple-choice questions in both English and Chinese versions, enabling an apples-to-apples comparison of VLMs' performance under a bilingual context. To summarize, MMBench is a systematically designed objective benchmark for …

Multimodal Large Language Models (MLLMs) excel in generating responses based on visual inputs. However, they often suffer from a bias towards generating responses similar to their pretraining corpus, overshadowing the importance of visual information. We treat this bias as a "preference" for pretraining statistics, which hinders the model's grounding in visual input. To mitigate this issue, we propose Bootstrapped Preference Optimization (BPO), which conducts preference learning with datasets containing negative responses bootstrapped from the model itself. Specifically, we propose the following two strategies: 1) using distorted image inputs to the MLLM for eliciting responses that contain signified pretraining bias; 2) leveraging text-based LLM to explicitly inject erroneous but common elements into the original response. Those undesirable responses are paired with original annotated responses from the datasets to construct the preference dataset, which is subsequently utilized to perform preference learning. Our approach effectively suppresses pretrained LLM bias, enabling enhanced grounding in visual inputs. Extensive experimentation demonstrates significant performance improvements across multiple benchmarks, advancing the state-of-the-art in multimodal conversational systems.

Dance, as an art form, fundamentally hinges on the precise synchronization with musical beats. However, achieving aesthetically pleasing dance sequences from music is challenging, with existing methods often falling short in controllability and beat alignment. To address these shortcomings, this paper introduces Beat-It, a novel framework for beat-specific, key pose-guided dance generation. Unlike prior approaches, Beat-It uniquely integrates explicit beat awareness and key pose guidance, effectively resolving two main issues: the misalignment of generated dance motions with musical beats, and the inability to map key poses to specific beats, critical for practical choreography. Our approach disentangles beat conditions from music using a nearest beat distance representation and employs a hierarchical multi-condition fusion mechanism. This mechanism seamlessly integrates key poses, beats, and music features, mitigating condition conflicts and offering rich, multi-conditioned guidance for dance generation. Additionally, a specially designed beat alignment loss ensures the generated dance movements remain in sync with the designated beats. Extensive experiments confirm Beat-It's superiority over existing state-of-the-art methods in terms of beat alignment and motion controllability. Qualitative results of our method can be found on our anonymous website.

The objective of this paper is to develop a functional system for translating spoken languages into sign languages, referred to as Spoken2Sign translation. The Spoken2Sign task is orthogonal and complementary to traditional sign language to spoken language (Sign2Spoken) translation. To enable Spoken2Sign translation, we present a simple baseline consisting of three steps: 1) creating a gloss-video dictionary using existing Sign2Spoken benchmarks; 2) estimating a 3D sign for each sign video in the dictionary; 3) training a Spoken2Sign model, which is composed of a Text2Gloss translator, a sign connector, and a rendering module, with the aid of the yielded gloss-3D sign dictionary. The translation results are then displayed through a sign avatar. As far as we know, we are the first to present the Spoken2Sign task in an output format of 3D signs. In addition to its capability of Spoken2Sign translation, we also demonstrate that two by-products of our approach—3D keypoint augmentation and multi-view understanding—can assist in keypoint-based sign language understanding. Code and models will be released to facilitate future research.

In this study, we identify the inefficient attention phenomena in Large Vision-Language Models (LVLMs), notably within prominent models like LLaVA-1.5, QwenVL-Chat and Video-LLaVA. We find out that the attention computation over visual tokens is of extreme inefficiency in the deep layers of popular LVLMs, suggesting a need for a sparser approach compared to textual data handling. To this end, we introduce FastV, a versatile plug-and-play method designed to optimize computational efficiency by learning adaptive attention patterns in early layers and pruning visual tokens in subsequent ones. Our evaluations demonstrate FastV's ability to dramatically reduce computational costs (e.g., a 45\% reduction in FLOPs for LLaVA-1.5-13B) without sacrificing performance in a wide range of image and video understanding tasks. The computational efficiency and performance trade-off of FastV are highly customizable and pareto-efficient. It can compress the FLOPs of a 13B-parameter model to achieve a lower budget than that of a 7B-parameter model, while still maintaining superior performance. We believe FastV has practical values for deployment of LVLMs in edge devices and commercial models. Code will be released upon acceptance.

This paper addresses a significant limitation that prevents Contrastive Language-Image Pretrained Models (CLIP) from achieving optimal performance on downstream image classification tasks. The key problem with CLIP-style zero-shot classification is that it requires domain-specific context in the form of prompts to better align the class descriptions to the downstream data distribution. In particular, prompts for vision-language models are domain-level texts (e.g., a centered satellite image of ...'') which, together with the class names, are fed into the text encoder to provide more context for the downstream dataset. These prompts are typically manually tuned, which is time consuming and often sub-optimal. To overcome this bottleneck, this paper proposes uCAP, a method to automatically learn domain-specific prompts/contexts using only unlabeled in-domain images. We achieve this by modeling the generation of images given the class names and a domain-specific prompt with an unsupervised likelihood distribution, and then performing inference of the prompts. We validate the proposed method across various models and datasets, showing that uCAP consistently outperforms manually tuned prompts and related baselines on the evaluated datasets: ImageNet, CIFAR-10, CIFAR-100, OxfordPets (up to 2\%), SUN397 (up to 5\%), and Caltech101 (up to 3\%).

Vision-language models (VLMs) are typically composed of a vision encoder, e.g. CLIP, and a language model (LM) that interprets the encoded features to solve downstream tasks. Despite remarkable progress, VLMs are subject to several shortcomings due to the limited capabilities of vision encoders, e.g. blindness'' to certain image features, visual hallucination, etc. To address these issues, we study broadening of the visual encoding capabilities of VLMs. We first comprehensively benchmark several vision encoders with different inductive biases for solving VLM tasks. We observe that there is no single encoding configuration that consistently achieves top performance across different tasks, and encoders with different biases can perform surprisingly similarly. Motivated by this, we introduce a method, named BRAVE, that consolidates features from multiple frozen encoders into a more versatile representation that can be directly fed as the input to a frozen LM. BRAVE achieves state-of-the-art performance on a broad range of captioning and VQA benchmarks and significantly reduces the aforementioned issues of VLMs, while requiring a smaller number of trainable parameters than existing methods and having a more compressed representation. Our results highlight the potential of incorporating different visual biases for a more broad and contextualized visual understanding of VLMs.

Recently, various methods have been proposed to address the inconsistency issue of DDIM inversion to enable image editing, such as EDICT [39] and Null-text inversion [23]. However, the above methods introduce considerable computational overhead. In this paper, we propose a new technique, named bi-directional integration approximation (BDIA), to perform exact diffusion inversion with negligible computational overhead. Suppose we would like to estimate the next diffusion state z{i-1} at timestep ti with the historical information (i, zi) and (i+1, z{i+1}). We first obtain the estimated Gaussian noise epsilon(zi, i), and then apply the DDIM update procedure twice for approximating the ODE integration over the next time-slot [ti, t{i-1}] in the forward manner and the previous time-slot [ti, t{t+1}] in the backward manner. The DDIM step for the previous time-slot is used to refine the integration approximation made earlier when computing zi. A nice property of BDIA-DDIM is that the update expression for z{i-1} is a linear combination of (z{i+1}, zi, epsilon(zi, i)). This allows for exact backward computation of z{i+1} given (zi, z_{i-1}), thus leading to exact diffusion inversion. We perform a convergence analysis for …

While personalized text-to-image generation has enabled the learning of a single concept from multiple images, a more practical yet challenging scenario involves learning multiple concepts within a single image. However, existing works tackling this scenario heavily rely on extensive human annotations. In this paper, we introduce a novel task named Unsupervised Concept Extraction (UCE) that considers an unsupervised setting without any human knowledge of the concepts. Given an image that contains multiple concepts, the task aims to extract and recreate individual concepts solely relying on the existing knowledge from pretrained diffusion models. To achieve this, we present ConceptExpress that tackles UCE by unleashing the inherent capabilities of pretrained diffusion models in two aspects. Specifically, a concept localization approach automatically locates and disentangles salient concepts by leveraging spatial correspondence from diffusion self-attention; and based on the lookup association between a concept and a conceptual token, a concept-wise optimization process learns discriminative tokens that represent each individual concept. Finally, we establish an evaluation protocol tailored for the UCE task. Extensive experiments demonstrate that ConceptExpress is a promising solution to the UCE task.

Recent developments in diffusion models have advanced conditioned image generation, yet they struggle with reconstructing out-of-distribution (OOD) images, such as unseen tumors, causing 'image hallucination' and risking misdiagnosis. We hypothesize such hallucinations result from local OOD regions in the conditional images. By partitioning the OOD region and conducting separate generations, hallucinations can be alleviated, and we verify this with motivational studies in several applications. From this, we propose a training-free diffusion framework that reduces hallucination by performing multiple \textit{Local Diffusion} processes. Our approach involves OOD estimation followed by two diffusion modules: a 'branching' module for local image generations from OOD estimations, and a 'fusion' module to integrate these predictions into a full image cohesively. These modules adapt to each testing dataset by updating an auxiliary classifier. Our evaluation shows our method improves baseline models quantitatively and qualitatively across different datasets. It also works well with various pre-trained diffusion models as a plug-and-play option.

We introduce Adversarial Diffusion Distillation (ADD), a novel training approach that efficiently samples large-scale foundational image diffusion models in just 1--4 steps while maintaining high image quality. We use score distillation to leverage large-scale off-the-shelf image diffusion models as a teacher signal in combination with an adversarial loss to ensure high image fidelity even in the low-step regime of one or two sampling steps. Our analyses show that our model clearly outperforms existing few-step methods (GANs, Latent Consistency Models) in a single step and reaches the performance of state-of-the-art diffusion models (SDXL) in only four steps. ADD is the first method to unlock single-step, real-time image synthesis with foundation models.

Diffusion models have shown remarkable results in generating 2D images and small-scale 3D objects. However, their application to the synthesis of large-scale 3D scenes has been rarely explored. This is mainly due to the inherent complexity and bulky size of 3D scenery data, particularly outdoor scenes, and the limited availability of comprehensive real-world datasets, which makes training a stable scene diffusion model challenging. In this work, we explore how to effectively generate large-scale 3D scenes using the coarse-to-fine paradigm. We introduce a framework, the Pyramid Discrete Diffusion model (PDD), which employs scale-varied diffusion models to progressively generate high-quality outdoor scenes. Experimental results of PDD demonstrate our successful exploration in generating 3D scenes both unconditionally and conditionally. We further showcase the data compatibility of the PDD model, due to its multi-scale architecture: a PDD model trained on one dataset can be easily fine-tuned with another dataset. The source codes and trained models will be made available to the public.

Humans naturally build mental models of object interactions and dynamics, allowing them to imagine how their surroundings will change if they take a certain action. While generative models today have shown impressive results on generating/editing images unconditionally or conditioned on text, current methods do not provide the ability to perform fine-grained object manipulation conditioned on actions, an important tool for world modeling and action planning. Therefore, we propose learning to model interactions through a novel form of visual conditioning: hands. Hands are a natural way to specify control through actions such as grasping, pulling, pushing, etc. Given an input image and a representation of a hand interacting with the scene, our approach, CoSHAND, synthesizes a depiction of what the scene would look like after the interaction has occurred. We show that CoSHAND is able to recover the dynamics of manipulation by learning from large amounts of unlabeled videos of human hands interacting with objects, and leveraging internet-scale latent diffusion model priors. The model demonstrates strong capabilities on a variety of actions and object types beyond the dataset, and the ability to generate multiple possible futures depending on the actions performed. CoSHAND is also able to generalize zero-shot to tasks where …

Federated Class Continual Learning (FCCL) merges the challenges of distributed client learning with the need for seamless adaptation to new classes without forgetting old ones. The key challenge in FCCL is catastrophic forgetting, an issue that has been explored to some extent in Continual Learning (CL). However, due to privacy preservation requirements, some conventional methods, such as experience replay, are not directly applicable to FCCL. Existing FCCL methods mitigate forgetting by generating historical data through federated training of GANs or data-free knowledge distillation. However, these approaches often suffer from unstable training of generators or low-quality generated data, limiting their guidance for the model. To address this challenge, we propose a novel method of data replay based on diffusion models. Instead of training a diffusion model, we employ a pre-trained conditional diffusion model to reverse-engineer each category, searching the corresponding input conditions for each category within the model's input space, significantly reducing computational resources and time consumption while ensuring effective generation. Furthermore, we enhance the classifier's domain generalization ability on generated and real data through contrastive learning, indirectly improving the representational capability of generated data for real data. Extensive experiments demonstrate that our method significantly outperforms existing baselines.

Omnidirectional images (ODIs) are commonly used in real-world visual tasks, and high-resolution ODIs help improve the performance of related visual tasks. Most existing super-resolution methods for ODIs use end-to-end learning strategies, resulting in inferior realness of generated images and a lack of effective out-of-domain generalization capabilities in training methods. Image generation methods represented by diffusion model provide strong priors for visual tasks and have been proven to be effectively applied to image restoration tasks. Leveraging the image priors of the Stable Diffusion (SD) model, we achieve omnidirectional image super-resolution with both fidelity and realness, dubbed as OmniSSR. Firstly, we transform the equirectangular projection (ERP) images into tangent projection (TP) images, whose distribution approximates the planar image domain. Then, we use SD to iteratively sample initial high-resolution results. At each denoising iteration, we further correct and update the initial results using the proposed Octadecaplex Tangent Information Interaction (OTII) and Gradient Decomposition (GD) technique to ensure better consistency. Finally, the TP images are transformed back to obtain the final high-resolution results. Our method is zero-shot, requiring no training or fine-tuning. Experiments of our method on two benchmark datasets demonstrate the effectiveness of our proposed method.

Action Quality Assessment (AQA) evaluates diverse skills but models struggle with non-stationary data. We propose Continual AQA (CAQA) to refine models using sparse new data. Feature replay preserves memory without storing raw inputs. However, the misalignment between static old features and the dynamically changing feature manifold causes severe catastrophic forgetting. To address this novel problem, we propose Manifold-Aligned Graph Regularization (MAGR), which first aligns deviated old features to the current feature manifold, ensuring representation consistency. It then constructs a graph jointly arranging old and new features aligned with quality scores. Experiments show MAGR outperforms recent strong baselines with up to 6.56%, 5.66%, 15.64%, and 9.05% correlation gains on the MTL-AQA, FineDiving, UNLV-Dive, and JDM-MSA split datasets, respectively. This validates MAGR for continual assessment challenges arising from non-stationary skill variations.

Compositional actions consist of dynamic (verbs) and static (objects) concepts. Humans can easily recognize unseen compositions using the learned concepts. For machines, solving such a problem requires a model to recognize unseen actions composed of previously observed verbs and objects, thus requiring, so-called, compositional generalization ability. To facilitate this research, we propose a novel Zero-Shot Compositional Action Recognition (ZS-CAR) task. For evaluating the task, we construct a new benchmark, Something-composition (Sth-com), based on the widely used Something-Something V2 dataset. We also propose a novel Component-to-Composition (C2C) learning method to solve the new ZS-CAR task. C2C includes an independent component learning module and a composition inference module. Last, we devise an enhanced training strategy to address the challenges of component variation between seen and unseen compositions and to handle the subtle balance between learning seen and unseen actions. The experimental results demonstrate that the proposed framework significantly surpasses the existing compositional generalization methods and sets a new state-of-the-art. The new Sth-com benchmark and code are available at https://anonymous.4open.science/r/C2C_anonymous-51F1.

Goal-oriented planning, or anticipating a series of actions that transition an agent from its current state to a predefined objective, is crucial for developing intelligent assistants aiding users in daily procedural tasks. The problem presents significant challenges due to the need for comprehensive knowledge of temporal and hierarchical task structures, as well as strong capabilities in reasoning and planning. To achieve this, prior work typically relies on extensive training on the target dataset, which often results in significant dataset bias and a lack of generalization to unseen tasks. In this work, we introduce VidAssist, an integrated framework designed for zero/few-shot goal-oriented planning in instructional videos. VidAssist leverages large language models (LLMs) as both the knowledge base and the assessment tool for generating and evaluating action plans, thus overcoming the challenges of acquiring procedural knowledge from small-scale, low-diversity datasets. Moreover, VidAssist employs a breadth-first search algorithm for optimal plan generation, in which a composite of value functions designed for goal-oriented planning are utilized to assess the predicted actions at each step. Extensive experiments demonstrate that VidAssist offers a unified framework for different goal-oriented planning setups, e.g., visual planning for assistance (VPA) and procedural planning (PP) and achieves remarkable performance in zero-shot …

The unprecedented surge in video data production in recent years necessitates efficient tools for extracting meaningful frames from videos for downstream tasks. Long-term temporal reasoning is a key desideratum for frame retrieval systems. While state-of-the-art foundation models, like VideoLLaMA and ViCLIP, are proficient in short-term semantic understanding, they surprisingly fail at long-term reasoning across frames. A key reason for their failure is that they intertwine per-frame perception and temporal reasoning into a single deep network. Hence, decoupling but co-designing semantic understanding and temporal reasoning is essential for efficient scene identification. We propose a system that leverages vision-language models for semantic understanding of individual frames but effectively reasons about the long-term evolution of events using state machines and temporal logic (TL) formulae that inherently capture memory. Our TL-based reasoning improves the F1 score of complex event identification by 9-15% compared to benchmarks that use GPT4 for reasoning on state-of-the-art self-driving datasets such as Waymo and NuScenes.

Video recognition models often learn scene-biased action representation due to the spurious correlation between actions and scenes in the training data. Such models show poor performance when the test data consists of videos with unseen action-scene combinations. Although Scene-debiased action recognition models might address the issue, they often overlook valuable scene information in the data. To address this challenge, we propose to learn Disentangled VIdeo representations of Action and Scene (DEVIAS), for more holistic video understanding. We propose an encoder-decoder architecture to learn disentangled action and scene representations with a single model. The architecture consists of a disentangling encoder (DE), an action mask decoder (AMD), and a prediction head. The key to achieving the disentanglement is employing both DE and AMD during training time. The DE uses the slot attention mechanism to learn disentangled action and scene representations. For further disentanglement, an AMD learns to predict action masks, given an action slot. With the resulting disentangled representations, we can achieve robust performance across diverse scenarios, including both seen and unseen action-scene combinations. We rigorously validate the proposed method on the UCF-101, Kinetics-400, and HVU datasets for the seen, and the SCUBA, HAT, and HVU datasets for unseen action-scene combination scenarios. …

Temporal video alignment aims to synchronize the key events like object interactions or action phase transitions in two videos. Such methods could benefit various video editing, processing, and understanding tasks. However, existing approaches operate under the restrictive assumption that a suitable video pair for alignment is given, significantly limiting their broader applicability. To address this, we re-pose temporal alignment as a search problem and introduce the task of Alignable Video Retrieval (AVR). Given a query video, our approach can identify well-alignable videos from a large collection of clips and temporally synchronize them to the query. To achieve this, we make three key contributions: 1) we introduce DRAQ, a video alignability indicator to identify and re-rank the best alignable video from a set of candidates; 2) we propose an effective and generalizable frame-level video feature design to improve the alignment performance of several off-the-shelf feature representations, and 3) we propose a novel benchmark and evaluation protocol for AVR using cycle-consistency metrics. Our experiments on 3 datasets, including large-scale Kinetics700, demonstrate the effectiveness of our approach in identifying alignable video pairs from diverse datasets.

Spatio-temporal video grounding aims to localize the spatio-temporal tube in a video according to the given language query. To eliminate the annotation costs, we make a first exploration to tackle spatio-temporal video grounding in a zero-shot manner. Our method dispenses with the need for any training videos or annotations; instead, it localizes the target object by leveraging large visual-language models and optimizing within the video and text query during the test time. To enable spatio-temporal comprehension, we introduce a multimodal modulation that integrates the spatio-temporal context into both visual and textual representation. On the visual side, we devise a context-based visual modulation that amplifies the visual representation by propagation and aggregation of the contextual semantics. Concurrently, on the textual front, we propose a prototype-based textual modulation to refine the textual representations using visual prototypes, effectively mitigating the cross-modal discrepancy. In addition, to overcome the interleaved spatio-temporal dilemma, we propose an expectation maximization (EM) framework to optimize the process of temporal relevance estimation and spatial region identification in an alternating way. Comprehensive experiments validate that our zero-shot approach achieves superior performance in comparison to several state-of-the-art methods with stronger supervision. We will make our code publicly accessible online.

We present a method to build animatable dog avatars from monocular videos. This is challenging as animals display a range of (unpredictable) non-rigid movements and have a variety of appearance details (e.g., fur, spots, tails). We develop an approach that links the video frames via a 4D solution that jointly solves for animal's pose variation, and its appearance (in a canonical pose). To this end, we significantly improve the quality of template-based shape fitting by endowing the SMAL parametric model with Continuous Surface Embeddings (CSE), which brings image-to-mesh reprojection constaints that are denser, and thus stronger, than the previously used sparse semantic keypoint correspondences. To model appearance, we propose an implicit duplex-mesh texture that is defined in the canonical pose, but can be deformed using SMAL pose coefficients and later rendered to enforce a photometric compatibility with the input video frames. On the challenging CoP3D and APTv2 datasets, we demonstrate superior results (both in terms of pose estimates and predicted appearance) to existing template-free (RAC) and template-based approaches (BARC, BITE)

Empowered by Large Language Models (LLMs), recent advancements in VideoLLMs have driven progress in various video understanding tasks. These models encode video representations through pooling or query aggregation over a vast amount of visual tokens, making computational and memory costs affordable. Despite successfully providing an overall comprehension of video content, existing VideoLLMs still face challenges in achieving detailed understanding in videos due to overlooking local information in long-term videos. To tackle this challenge, we introduce LongVLM, a straightforward yet powerful VideoLLM for long video understanding, building upon the observation that long videos often consist of sequential key events, complex actions, and camera movements. Our approach proposes to decompose long videos into multiple short-term segments and encode local features for each local segment via a hierarchical token merging module. These features are concatenated in temporal order to maintain the storyline across sequential short-term segments. Additionally, we propose to integrate global semantics into each local feature to enhance context understanding. In this way, we encode video representations that incorporate both local and global information, enabling the LLM to generate comprehensive responses for long-term videos. Experimental results on the VideoChatGPT benchmark and zero-shot video question-answering datasets demonstrate the superior capabilities of our model …

Our objective is to discover and localize monotonic temporal changes in a sequence of images. To achieve this, we exploit a simple proxy task of ordering a shuffled image sequence, with `time' serving as a supervisory signal since only changes that are monotonic with time can give rise to the correct ordering. We also introduce a flexible transformer-based model for general-purpose ordering of image sequences of arbitrary length with built-in attribution maps. After training, the model successfully discovers and localizes monotonic changes while ignoring cyclic and stochastic ones. We demonstrate applications of the model in multiple video settings covering different scene and object types, discovering both object-level and environmental changes in unseen sequences. We also demonstrate that the attention-based attribution maps function as effective prompts for segmenting the changing regions, and that the learned representations can be used for downstream applications. Finally, we show that the model achieves the state of the art on standard benchmarks for ordering a set of images.

Multimodal Large Language Models (MLLMs) excel in generating responses based on visual inputs. However, they often suffer from a bias towards generating responses similar to their pretraining corpus, overshadowing the importance of visual information. We treat this bias as a "preference" for pretraining statistics, which hinders the model's grounding in visual input. To mitigate this issue, we propose Bootstrapped Preference Optimization (BPO), which conducts preference learning with datasets containing negative responses bootstrapped from the model itself. Specifically, we propose the following two strategies: 1) using distorted image inputs to the MLLM for eliciting responses that contain signified pretraining bias; 2) leveraging text-based LLM to explicitly inject erroneous but common elements into the original response. Those undesirable responses are paired with original annotated responses from the datasets to construct the preference dataset, which is subsequently utilized to perform preference learning. Our approach effectively suppresses pretrained LLM bias, enabling enhanced grounding in visual inputs. Extensive experimentation demonstrates significant performance improvements across multiple benchmarks, advancing the state-of-the-art in multimodal conversational systems.

The objective of this paper is to develop a functional system for translating spoken languages into sign languages, referred to as Spoken2Sign translation. The Spoken2Sign task is orthogonal and complementary to traditional sign language to spoken language (Sign2Spoken) translation. To enable Spoken2Sign translation, we present a simple baseline consisting of three steps: 1) creating a gloss-video dictionary using existing Sign2Spoken benchmarks; 2) estimating a 3D sign for each sign video in the dictionary; 3) training a Spoken2Sign model, which is composed of a Text2Gloss translator, a sign connector, and a rendering module, with the aid of the yielded gloss-3D sign dictionary. The translation results are then displayed through a sign avatar. As far as we know, we are the first to present the Spoken2Sign task in an output format of 3D signs. In addition to its capability of Spoken2Sign translation, we also demonstrate that two by-products of our approach—3D keypoint augmentation and multi-view understanding—can assist in keypoint-based sign language understanding. Code and models will be released to facilitate future research.

Vision-Language Large Models (VLMs) recently become primary backbone of AI, due to the impressive performance. However, their expensive computation costs, i.e., throughput and delay, impede potentials in the real-world scenarios. To achieve acceleration for VLMs, most existing methods focus on the model perspective: pruning, distillation, quantization, but completely overlook the data-perspective redundancy. To fill the overlook, this paper pioneers the severity of data redundancy, and designs one plug-and-play Turbo module guided by information degree to prune inefficient tokens from visual or textual data. In pursuit of efficiency-performance trade-offs, information degree takes two crucial factors into consideration: mutual redundancy and semantic value. Concretely, the former evaluates data duplication between sequential tokens; while the latter evaluates each token by its contribution to the overall semantics. As a result, tokens with high information degree carry less redundancy and stronger semantics. For VLMs' calculation, Turbo works as a user-friendly plug-in that sorts data referring to information degree, utilizing only top-level ones to save costs. Its advantages are multifaceted, e.g., being generally compatible to various VLMs across understanding and generation, simple use without retraining and trivial engineering efforts. On multiple VLMs benchmarks, we fully experiment to reveal good acceleration of Turbo, under negligible performance drop.

Dance, as an art form, fundamentally hinges on the precise synchronization with musical beats. However, achieving aesthetically pleasing dance sequences from music is challenging, with existing methods often falling short in controllability and beat alignment. To address these shortcomings, this paper introduces Beat-It, a novel framework for beat-specific, key pose-guided dance generation. Unlike prior approaches, Beat-It uniquely integrates explicit beat awareness and key pose guidance, effectively resolving two main issues: the misalignment of generated dance motions with musical beats, and the inability to map key poses to specific beats, critical for practical choreography. Our approach disentangles beat conditions from music using a nearest beat distance representation and employs a hierarchical multi-condition fusion mechanism. This mechanism seamlessly integrates key poses, beats, and music features, mitigating condition conflicts and offering rich, multi-conditioned guidance for dance generation. Additionally, a specially designed beat alignment loss ensures the generated dance movements remain in sync with the designated beats. Extensive experiments confirm Beat-It's superiority over existing state-of-the-art methods in terms of beat alignment and motion controllability. Qualitative results of our method can be found on our anonymous website.

Vision-language models (VLMs) are typically composed of a vision encoder, e.g. CLIP, and a language model (LM) that interprets the encoded features to solve downstream tasks. Despite remarkable progress, VLMs are subject to several shortcomings due to the limited capabilities of vision encoders, e.g. blindness'' to certain image features, visual hallucination, etc. To address these issues, we study broadening of the visual encoding capabilities of VLMs. We first comprehensively benchmark several vision encoders with different inductive biases for solving VLM tasks. We observe that there is no single encoding configuration that consistently achieves top performance across different tasks, and encoders with different biases can perform surprisingly similarly. Motivated by this, we introduce a method, named BRAVE, that consolidates features from multiple frozen encoders into a more versatile representation that can be directly fed as the input to a frozen LM. BRAVE achieves state-of-the-art performance on a broad range of captioning and VQA benchmarks and significantly reduces the aforementioned issues of VLMs, while requiring a smaller number of trainable parameters than existing methods and having a more compressed representation. Our results highlight the potential of incorporating different visual biases for a more broad and contextualized visual understanding of VLMs.

Large vision-language models (VLMs) have recently achieved remarkable progress, exhibiting impressive multimodal perception and reasoning abilities. However, effectively evaluating these large VLMs remains a major challenge, hindering future development in this domain. Traditional benchmarks like VQAv2 or COCO Caption provide quantitative performance measurements but lack fine-grained ability assessment and robust evaluation metrics. Meanwhile, subjective benchmarks, such as OwlEval, offer comprehensive evaluations of a model's abilities by incorporating human labor, which is not scalable and may display significant bias. In response to these challenges, we propose MMBench, a bilingual benchmark for assessing the multi-modal capabilities of VLMs. MMBench methodically develops a comprehensive evaluation pipeline, primarily comprised of the following key features: 1. MMBench is meticulously curated with well-designed quality control schemes, surpassing existing similar benchmarks in terms of the number and variety of evaluation questions and abilities; 2. MMBench introduces a rigorous CircularEval strategy and incorporates large language models to convert free-form predictions into pre-defined choices, which helps to yield accurate evaluation results for models with limited instruction-following capabilities. 3. MMBench incorporates multiple-choice questions in both English and Chinese versions, enabling an apples-to-apples comparison of VLMs' performance under a bilingual context. To summarize, MMBench is a systematically designed objective benchmark for …

This paper addresses a significant limitation that prevents Contrastive Language-Image Pretrained Models (CLIP) from achieving optimal performance on downstream image classification tasks. The key problem with CLIP-style zero-shot classification is that it requires domain-specific context in the form of prompts to better align the class descriptions to the downstream data distribution. In particular, prompts for vision-language models are domain-level texts (e.g., a centered satellite image of ...'') which, together with the class names, are fed into the text encoder to provide more context for the downstream dataset. These prompts are typically manually tuned, which is time consuming and often sub-optimal. To overcome this bottleneck, this paper proposes uCAP, a method to automatically learn domain-specific prompts/contexts using only unlabeled in-domain images. We achieve this by modeling the generation of images given the class names and a domain-specific prompt with an unsupervised likelihood distribution, and then performing inference of the prompts. We validate the proposed method across various models and datasets, showing that uCAP consistently outperforms manually tuned prompts and related baselines on the evaluated datasets: ImageNet, CIFAR-10, CIFAR-100, OxfordPets (up to 2\%), SUN397 (up to 5\%), and Caltech101 (up to 3\%).

In this study, we identify the inefficient attention phenomena in Large Vision-Language Models (LVLMs), notably within prominent models like LLaVA-1.5, QwenVL-Chat and Video-LLaVA. We find out that the attention computation over visual tokens is of extreme inefficiency in the deep layers of popular LVLMs, suggesting a need for a sparser approach compared to textual data handling. To this end, we introduce FastV, a versatile plug-and-play method designed to optimize computational efficiency by learning adaptive attention patterns in early layers and pruning visual tokens in subsequent ones. Our evaluations demonstrate FastV's ability to dramatically reduce computational costs (e.g., a 45\% reduction in FLOPs for LLaVA-1.5-13B) without sacrificing performance in a wide range of image and video understanding tasks. The computational efficiency and performance trade-off of FastV are highly customizable and pareto-efficient. It can compress the FLOPs of a 13B-parameter model to achieve a lower budget than that of a 7B-parameter model, while still maintaining superior performance. We believe FastV has practical values for deployment of LVLMs in edge devices and commercial models. Code will be released upon acceptance.

This paper proposes a simple, yet effective framework, called GiT, simultaneously applicable for various vision tasks only with a vanilla ViT. Motivated by the universality of the Multi-layer Transformer architecture (e.g., GPT) widely used in large language models (LLMs), we seek to broaden its scope to serve as a powerful vision foundation model (VFM). However, unlike language modeling, visual tasks typically require specific modules, such as bounding box heads for detection and pixel decoders for segmentation, greatly hindering the application of powerful multi-layer transformers in the vision domain. To solve this, we design a universal language interface that empowers the successful auto-regressive decoding to adeptly unify various visual tasks, from image-level understanding (e.g., captioning), over sparse perception (e.g., detection), to dense prediction (e.g., segmentation). Based on the above designs, the entire model is composed solely of a ViT, without any specific additions, offering a remarkable architectural simplification. GiT is a multi-task visual model, jointly trained across five representative benchmarks without task-specific fine-tuning. Interestingly, our GiT builds a new benchmark in generalist performance, and fosters mutual enhancement across tasks, leading to significant improvements compared to isolated training. This reflects a similar impact observed in LLMs. Further enriching training with 27 datasets, …

Vision-Language Pre-training (VLP) models like CLIP have achieved remarkable success in computer vision and particularly demonstrated superior robustness to distribution shifts of 2D images. However, their robustness under 3D viewpoint variations is still limited, which can hinder the development for real-world applications. This paper successfully addresses this concern while keeping VLPs' original performance by breaking through two primary obstacles: 1) the scarcity of training data and 2) the suboptimal fine-tuning paradigms. To combat data scarcity, we build the Multi-View Caption (MVCap) dataset --- a comprehensive collection of over four million multi-view image-text pairs across more than 100K objects, providing more potential for VLP models to develop generalizable viewpoint-invariant representations. To address the limitations of existing paradigms in performance trade-offs and training efficiency, we design a novel fine-tuning framework named Omniview-Tuning (OVT). Specifically, OVT introduces a Cross-Viewpoint Alignment objective through a minimax-like optimization strategy, which effectively aligns representations of identical objects from diverse viewpoints without causing overfitting. Additionally, OVT fine-tunes VLP models in a parameter-efficient manner, leading to minimal computational cost. Extensive experiments on various VLP models with different architectures validate that OVT significantly improves the models' resilience to viewpoint shifts and keeps the original performance, establishing a pioneering standard for …

Recent years have witnessed considerable achievements in facial avatar reconstruction with neural volume rendering. Despite notable advancements, the reconstruction of complex and dynamic head movements from monocular videos still suffers from capturing and restoring fine-grained details. In this work, we propose a novel approach, named Tri^2-plane, for monocular photo-realistic volumetric head avatar reconstructions. Distinct from the existing works that rely on a single tri-plane deformation field for dynamic facial modeling, the proposed Tri^2-plane leverages the principle of feature pyramids and three top-to-down lateral connections tri-planes for details improvement. It samples and renders facial details at multiple scales, transitioning from the entire face to specific local regions and then to even more refined sub-regions. Moreover, we incorporate a camera-based geometry-aware sliding window method as an augmentation in training, which improves the robustness beyond the canonical space, with a particular improvement in cross-identity generation capabilities. Experimental outcomes indicate that the Tri^2-plane not only surpasses existing methodologies but also achieves superior performance across quantitative and qualitative assessments.

Despite progress in human motion capture, existing multi-view methods often face challenges in estimating the 3D pose and shape of multiple closely interacting people. This difficulty arises from reliance on accurate 2D joint estimations, which are hard to obtain due to occlusions and body contact when people are in close interaction. To address this, we propose a novel method leveraging the personalized implicit neural avatar of each individual as a prior, which significantly improves the robustness and precision of this challenging pose estimation task. Concretely, the avatars are efficiently reconstructed via layered volume rendering from sparse multi-view videos. The reconstructed avatar prior allows for direct optimization of 3D poses based on color and silhouette rendering loss, bypassing the issues associated with noisy 2D detections. To handle interpenetration, we propose a collision loss on the overlapping shape region of avatars to add penetration constraints. Moreover, both 3D poses and avatars are optimized in an alternating manner. Our experimental results demonstrate state-of-the-art performance on several public datasets.

The field of photorealistic 3D avatar reconstruction and generation has garnered significant attention in recent years; however, animating such avatars remains challenging. Recent advances in diffusion models have notably enhanced the capabilities of generative models in 2D animation. In this work, we directly utilize these models within the 3D domain to achieve controllable and high-fidelity 4D facial animation. By integrating the strengths of diffusion processes and geometric deep learning, we employ Graph Neural Networks (GNNs) as denoising diffusion models in a novel approach, formulating the diffusion process directly on the mesh space and enabling the generation of 3D facial expressions. This facilitates the generation of facial deformations through a mesh-diffusion-based model. Additionally, to ensure temporal coherence in our animations, we propose a consistent noise sampling method. Under a series of both quantitative and qualitative experiments, we showcase that the proposed method outperforms prior work in 4D expression synthesis by generating high-fidelity extreme expressions. Furthermore, we applied our method to textured 4D facial expression generation, implementing a straightforward extension that involves training on a large-scale textured 4D facial expression database.

The spike camera continuously records scene radiance with high-speed, high dynamic range, and low data redundancy properties, as a promising replacement for frame-based high-speed cameras. Previous methods for reconstructing color videos from monochromatic spikes are constrained in capturing full-temporal color information due to their reliance on compensating colors from low-speed RGB frames. Applying a Bayer-pattern color filter array to the spike sensor yields mosaicked chromatic spikes, which complicates noise distribution in high-speed conditions. By validating that the noise of short-term frames follows a zero-mean distribution, we leverage this hypothesis to develop a self-supervised denoising module trained exclusively on real-world data. Although noise is reduced in short-term frames, the long-term accumulation of incident photons is still necessary to construct HDR frames. Therefore, we introduce a progressive warping module to generate pseudo long-term exposure frames. This approach effectively mitigates motion blur artifacts in high-speed conditions. Integrating these modules forms a real-data-driven reconstruction method for mosaicked chromatic spikes. Extensive experiments conducted on both synthetic and real-world data demonstrate that our approach is effective in reconstructing 2000FPS color HDR videos with significantly reduced noise and motion blur compared to existing methods.

The integration of miniaturized spectrometers into mobile devices offers new avenues for image quality enhancement and facilitates novel downstream tasks. However, the broader application of spectral sensors in mobile photography is hindered by the inherent complexity of spectral images and the constraints of spectral imaging capabilities. To overcome these challenges, we propose a joint RGB-Spectral decomposition model guided enhancement framework, which consists of two steps: joint decomposition and priors-guided enhancement. Firstly, we leverage the complementarity between RGB and Low-resolution Multi-Spectral Images (Lr-MSI) to predict shading, reflectance, and material semantic priors. Subsequently, these priors are seamlessly integrated into the established HDRNet to promote dynamic range enhancement, color mapping, and grid expert learning, respectively. Additionally, we construct a high-quality Mobile-Spec dataset to support our research, and our experiments validate the effectiveness of Lr-MSI in the tone enhancement task. This work aims to establish a solid foundation for advancing spectral vision in mobile photography.

We introduce a simple yet effective approach for separating transmitted and reflected light. Our key insight is that the powerful novel view synthesis capabilities provided by modern inverse rendering methods (e.g.,~Gaussian splatting) allow one to perform flash/no-flash reflection separation using {\em unpaired measurements}---this relaxation dramatically simplifies image acquisition over conventional paired flash/no-flash reflection separation methods. Through extensive real-world experiments, we demonstrate our method, Flash-Splat, accurately reconstructs both transmitted and reflected scenes in 3D. Our method outperforms existing 3D reflection separation methods, which do not leverage illumination control, by a large margin.

Underwater scenes are challenging for computer vision methods due to color degradation caused by the water column and detrimental lighting effects such as caustic caused by sunlight refracting on a wavy surface. These challenges impede widespread use of computer vision tools that could aid in ecological surveying of underwater environments or in industrial applications. Existing algorithms for alleviating caustics and descattering the image to recover colors are often impractical to implement due to the need for ground-truth training data, the necessity for successful alignment of an image within a 3D scene, or other assumptions that are infeasible in practice. In this paper, we propose a solution to tackle those problems in underwater computer vision: our method is based on two neural networks: CausticsNet, for single-image caustics removal, and BackscatterNet, for backscatter removal. Both neural networks are trained using an objective formulated with the aid of self-supervised monocular SLAM on a collection of underwater videos. Thus, our method does not requires any ground-truth color images or caustics labels, and corrects images in real-time. We experimentally demonstrate the fidelity of our caustics removal method, performing similarly to state-of-the-art supervised methods, and show that the color restoration and caustics removal lead to better …

Novel-view synthesis based on visible light has been extensively studied. In comparison to visible light imaging, thermal infrared imaging offers the advantage of all-weather imaging and strong penetration, providing increased possibilities for reconstruction in nighttime and adverse weather scenarios. However, thermal infrared imaging is influenced by physical characteristics such as atmospheric transmission effects and thermal conduction, hindering the precise reconstruction of intricate details in thermal infrared scenes, manifesting as issues of floaters and indistinct edge features in synthesized images. To address these limitations, this paper introduces a physics-induced 3D Gaussian splatting method named Thermal3D-GS. Thermal3D-GS begins by modeling atmospheric transmission effects and thermal conduction in three-dimensional media using neural networks. Additionally, a temperature consistency constraint is incorporated into the optimization objective to enhance the reconstruction accuracy of thermal infrared images. Furthermore, to validate the effectiveness of our method, the first large-scale benchmark dataset for this field named Thermal Infrared Novel-view Synthesis Dataset (TI-NSD) is created. This dataset comprises 20 authentic thermal infrared video scenes, covering indoor, outdoor, and UAV(Unmanned Aerial Vehicle) scenarios, totaling 6,664 frames of thermal infrared image data. Based on this dataset, this paper experimentally verifies the effectiveness of Thermal3D-GS. The results indicate that our method outperforms the …

Implicit Neural Representation (INR) has become a popular method for representing visual signals (\eg, 2D images and 3D scenes), showing promising results in various downstream applications. Given its potential as a medium for visual signals, exploring the development of a neural blending method that utilizes INRs is a natural progression. Neural blending involves merging two INRs to create a new INR that encapsulates information from both original representations. A direct approach involves applying traditional image editing methods to the INR rendering process. However, this method often results in blending distortions, artifacts, and color shifts, primarily due to the discretization of the underlying discrete pixel grid and the introduction of boundary conditions for solving variational problems. To tackle this issue, we introduce the Neural Poisson Solver, a plug-and-play and universally applicable framework across different signal dimensions for blending visual signals represented by INRs. Our Neural Poisson Solver offers a variational problem-solving approach based on the continuous Poisson equation, which has demonstrated exceptional performance across various domains. Specifically, we propose a gradient-guided neural solver to represent the solution process of the variational problem, refining the target signal to achieve natural blending results. We also develop a Poisson equation-based loss and optimization scheme …

Typical inverse rendering methods focus on learning implicit neural scene representations by modeling the geometry, materials and illumination separately, which entails significant computations for optimization. In this work we design a Unified Voxelization framework for explicit learning of scene representations, dubbed UniVoxel, which allows for efficient modeling of the geometry, materials and illumination jointly, thereby accelerating the inverse rendering significantly. To be specific, we propose to encode a scene into a latent volumetric representation, based on which the geometry, materials and illumination can be readily learned via lightweight neural networks in a unified manner. Particularly, an essential design of UniVoxel is that we leverage local Spherical Gaussians to represent the incident light radiance, which enables the seamless integration of modeling illumination into the unified voxelization framework. Such novel design enables our UniVoxel to model the joint effects of direct lighting, indirect lighting and light visibility efficiently without expensive multi-bounce ray tracing. Extensive experiments on multiple benchmarks covering diverse scenes demonstrate that UniVoxel boosts the optimization efficiency significantly compared to other methods, reducing the per-scene training time from hours to 18 minutes, while achieving favorable reconstruction quality. Code will be released.

Existing neural radiance field-based methods can achieve real-time rendering of small scenes on the web platform. However, extending these methods to large-scale scenes still poses significant challenges due to limited resources in computation, memory, and bandwidth. In this paper, we propose City-on-Web, the first method for real-time rendering of large-scale scenes on the web. We propose a block-based volume rendering method to guarantee 3D consistency and correct occlusion between blocks, and introduce a Level-of-Detail strategy combined with dynamic loading/unloading of resources to significantly reduce memory demands. Our system achieves real-time rendering of large-scale scenes at approximately 32FPS with RTX 3060 GPU on the web and maintains rendering quality comparable to the current state-of-the-art novel view synthesis methods.

Novel view synthesis with sparse inputs poses great challenges to Neural Radiance Field (NeRF). Recent works demonstrate that the frequency regularization of Positional Encoding (PE) can achieve promising results for this task. In this work, we reveal that there exists an inconsistency between the frequency regularization of PE and rendering loss. This prevents few-shot NeRF from synthesizing higher-quality novel views. To mitigate this inconsistency, we propose Adaptive Rendering loss regularization for few-shot NeRF, dubbed AR-NeRF. Specifically, we present a two-phase rendering supervision and an adaptive rendering loss weight learning strategy to align the frequency relationship between PE and 2D-pixel supervision. In this way, AR-NeRF can learn global structures better in the early training phase and adaptively learn local details throughout the training process. Extensive experiments show that our AR-NeRF achieves state-of-the-art performance on different datasets, including object-level and complex scenes. Our code will be released upon publication.

While neural rendering has demonstrated impressive capabilities in 3D scene reconstruction and novel view synthesis, it heavily relies on high-quality sharp images and accurate camera poses. Numerous approaches have been proposed to train Neural Radiance Fields (NeRF) with motion-blurred images, commonly encountered in real-world scenarios such as low-light or long-exposure conditions. However, the implicit representation of NeRF struggles to accurately recover intricate details from severely motion-blurred images and cannot achieve real-time rendering. In contrast, recent advancements in 3D Gaussian Splatting achieve high-quality 3D scene reconstruction and real-time rendering by explicitly optimizing point clouds as Gaussian spheres. In this paper, we introduce a novel approach, named BAD-Gaussians (Bundle Adjusted Deblur Gaussian Splatting), which leverages explicit Gaussian representation and handles severe motion-blurred images with inaccurate camera poses to achieve high-quality scene reconstruction. Our method models the physical image formation process of motion-blurred images and jointly learns the parameters of Gaussians while recovering camera motion trajectories during exposure time. In our experiments, we demonstrate that BAD-Gaussians not only achieves superior rendering quality compared to previous state-of-the-art deblur neural rendering methods on both synthetic and real datasets but also enables real-time rendering capabilities.

Recent progress in neural rendering have brought forth pioneering methods, such as NeRF and Gaussian Splatting, which revolutionize view rendering across various domains like AR/VR, gaming, and content creation. While these methods excel at interpolating within the training data, the challenge of generalizing to new scenes and objects from very sparse views persists. Specifically, modeling 3D humans from sparse views presents formidable hurdles due to the inherent complexity of human geometry, resulting in inaccurate reconstructions of geometry and textures. To tackle this challenge, this paper leverages recent advancements in Gaussian splatting and introduces a new method to learn generalizable human Gaussians that allows photorealistic and accurate view-rendering of a new human subject from a limited set of sparse views in a feed-forward manner. A pivotal innovation of our approach involves reformulating the learning of 3D Gaussian parameters into a regression process defined on the 2D UV space of a human template, which allows leveraging the strong geometry prior and the advantages of 2D convolutions. Our method outperforms recent methods on both within-dataset generalization as well as cross-dataset generalization settings.

This paper tackles the simultaneous optimization of pose and Neural Radiance Fields (NeRF). Departing from the conventional practice of using explicit global representations for camera pose, we introduce the Invertible Neural Warp, a novel overparameterized representation for camera pose. Specifically, we establish that invertibility is a crucial property when using Multi-layer Perceptrons (MLPs) for ray transformation. This work proposes utilizing an Invertible Neural Network, coupled with a geometry-informed constraint to achieve enhanced optimization convergence. Experiments on the LLFF and DTU datasets demonstrate the effectiveness of our approach in terms of pose estimation and high-fidelity reconstruction compared to existing well-established approaches.

Neural implicit surface reconstruction has achieved remarkable progress recently. Despite resorting to complex radiance modeling, state-of-the-art methods still struggle with textureless and specular surfaces. Different from RGB images, polarization images can provide direct constraints on the azimuth angles of the surface normals. In this paper, we present PISR, a novel method that utilizes a geometrically accurate polarimetric loss to refine shape independently of appearance. In addition, PISR smooths surface normals in image space to eliminate severe shape distortions and leverages the hash-grid-based neural signed distance function to accelerate the reconstruction. Experimental results demonstrate that PISR achieves higher accuracy and robustness, with an L1 chamfer distance of 0.5 mm and an F-score of 99.5% at 1mm, while converging 4~ 30x faster than previous polarimetric surface reconstruction methods. The source code and dataset will be released upon acceptance of this paper.

Recent advancements in Neural Surface Reconstruction (NSR) have significantly improved multi-view reconstruction when coupled with volume rendering. However, relying solely on photometric consistency in image space falls short in addressing complexities posed by real-world data, including occlusions and non-Lambertian surfaces. To tackle these challenges, we propose an investigation into feature-level consistent loss, aiming to harness valuable feature priors from diverse pretext visual tasks and overcome current limitations. It is crucial to note the existing gap in determining the most effective pretext visual task for enhancing NSR. In this study, we comprehensively explore multi-view feature priors from seven pretext visual tasks, comprising thirteen methods. Our main goal is to strengthen NSR training by considering a wide range of possibilities. Additionally, we examine the impact of varying feature resolutions and evaluate both pixel-wise and patch-wise consistent losses, providing insights into effective strategies for improving NSR performance. By incorporating pre-trained representations from MVSFormer and QuadTree, our approach can generate variations of MVS-NeuS and Match-NeuS, respectively. Our results, analyzed on DTU and EPFL datasets, reveal that feature priors from image matching and multi-view stereo outperform other pretext tasks. Moreover, we discover that extending patch-wise photometric consistency to the feature level surpasses the performance of …

3D surface reconstruction from images is essential for numerous applications. Recently, Neural Radiance Fields (NeRFs) have emerged as a promising framework for 3D modeling. However, NeRFs require accurate camera poses as input, and existing methods struggle to handle significantly noisy pose estimates (i.e., outliers), which are commonly encountered in real-world scenarios. To tackle this challenge, we present a novel approach that optimizes radiance fields with scene graphs to mitigate the influence of outlier poses. Our method incorporates an adaptive inlier-outlier confidence estimation scheme based on scene graphs, emphasizing images of high compatibility with the neighborhood and consistency in the rendering quality. We also introduce an effective intersection-over-union (IoU) loss to optimize the camera pose and surface geometry, together with a coarse-to-fine strategy to facilitate the training. Furthermore, we propose a new dataset containing typical outlier poses for a detailed evaluation. Experimental results on various datasets consistently demonstrate the effectiveness and superiority of our method over existing approaches, showcasing its robustness in handling outliers and producing high-quality 3D reconstructions.

Novel view synthesis from unconstrained in-the-wild images remains a meaningful but challenging task. The photometric variation and transient occluders in those unconstrained images make it difficult to reconstruct the original scene accurately. Previous approaches tackle the problem by introducing a global appearance feature in Neural Radiance Fields (NeRF). However, in the real world, the unique appearance of each tiny point in a scene is determined by its independent intrinsic material attributes and the varying environmental impacts it receives. Inspired by this fact, we propose Gaussian in the wild (GS-W), a method that uses 3D Gaussian points to reconstruct the scene and introduces separated intrinsic and dynamic appearance feature for each point, capturing the unchanged scene appearance along with dynamic variation like illumination and weather. Additionally, an adaptive sampling strategy is presented to allow each Gaussian point to focus on the local and detailed information more effectively. We also reduce the impact of transient occluders using a 2D visibility map. More experiments have demonstrated better reconstruction quality and details of GS-W compared to previous methods, with a 1000 times increase in rendering speed.

The use of 3D Gaussians as representation of radiance fields has enabled high quality novel view synthesis at real-time rendering speed. However, the choice of optimising the outgoing radiance of each Gaussian independently as spherical harmonics results in unsatisfactory view dependent effects. In response to these limitations, our work, Factorised Tensorial Illumination for 3D Gaussian Splatting or `3iGS' improves upon 3D Gaussian Splatting (3DGS) rendering quality. Instead of optimising a single outgoing radiance parameter, 3iGS enhances 3DGS view dependent effects by expressing the outgoing radiance as a function of a local illumination field and Bidirectional Reflectance Distribution Function (BRDF) features. We optimise a continuous incident illumination field through a Tensorial Factorisation representation, while separately fine-tuning the BRDF features of each 3D Gaussian relative to this illumination field. Our methodology significantly enhances the rendering quality of specular view-dependent effects of 3DGS, while maintaining rapid training and rendering speeds. Code will be released.

The rapid growth of 3D Gaussian Splatting (3DGS) has revolutionized neural rendering, enabling real-time production of high-quality renderings. However, the previous 3DGS-based methods have limitations in urban scenes due to reliance on initial Structure-from-Motion(SfM) points and difficulties in rendering distant, sky and low-texture areas. To overcome these challenges, we propose a hybrid optimization method named HO-Gaussian, which combines a grid-based volume with the 3DGS pipeline. HO-Gaussian eliminates the dependency on SfM point initialization, allowing for rendering of urban scenes, and incorporates the Point Densitification to enhance rendering quality in problematic regions during training. Furthermore, we introduce Gaussian Direction Encoding as an alternative for spherical harmonics in the rendering pipeline, which enables view-dependent color representation. To adapt the 3DGS method for multi-camera systems, we introduce neural warping to enhance object consistency across different cameras. Experimental results on widely used autonomous driving datasets demonstrate that HO-Gaussian achieves photo-realistic rendering in real-time on multi-camera urban datasets.

During the Gaussian Splatting optimization process, the scene's geometry can gradually deteriorate if its structure is not deliberately preserved, especially in non-textured regions such as walls, ceilings, and furniture surfaces. This degradation significantly affects the rendering quality of novel views that deviate significantly from the viewpoints in the training data. To mitigate this issue, we propose a novel approach called GeoGaussian. Based on the smoothly connected areas observed from point clouds, this method introduces a novel pipeline to initialize thin Gaussians aligned with the surfaces, where the characteristic can be transfered to new generations through a carefully designed densification strategy. Finally, the pipeline ensures that the scene's geometry and texture are maintained through constrained optimization processes with explicit geometry constraints. Benefiting from the proposed architecture, the generative ability of 3D Gaussians is enhanced, especially in structured regions. Our proposed pipeline achieves state-of-the-art performance in novel view synthesis and geometric reconstruction, as evaluated qualitatively and quantitatively on public datasets.

Recently, 3D Gaussian splatting (3D-GS) has gained popularity in novel-view scene synthesis. It addresses the challenges of lengthy training times and slow rendering speeds associated with Neural Radiance Fields (NeRFs). Through rapid, differentiable rasterization of 3D Gaussians, 3D-GS achieves real-time rendering and accelerated training. They, however, demand substantial memory resources for both training and storage, as they require millions of Gaussians in their point cloud representation for each scene. We present a technique utilizing quantized embeddings to significantly reduce per-point memory storage requirements and a coarse-to-fine training strategy for a faster and more stable optimization of the Gaussian point clouds. Our approach develops a pruning stage which results in scene representations with fewer Gaussians, leading to faster training times and rendering speeds for real-time rendering of high resolution scenes. We reduce storage memory by more than an order of magnitude all while preserving the reconstruction quality. We validate the effectiveness of our approach on a variety of datasets and scenes preserving the visual quality while consuming 10-20x less memory and faster training/inference speed.

3D Gaussian Splatting (3DGS) has become an emerging technique with remarkable potential in 3D representation and image rendering. However, the substantial storage overhead of 3DGS significantly impedes its practical applications. In this work, we formulate the compact 3D Gaussian learning as an end-to-end Rate-Distortion Optimization (RDO) problem and propose RDO-Gaussian that can achieve flexible and continuous rate control. RDO-Gaussian addresses two main issues that exist in current schemes: 1) Different from prior endeavors that minimize the rate under the fixed distortion, we introduce dynamic pruning and entropy-constrained vector quantization (ECVQ) that optimize the rate and distortion at the same time. 2) Previous works treat the colors of each Gaussian equally, while we model the colors of different regions and materials with learnable numbers of parameters. We verify our method on both real and synthetic scenes, showcasing that RDO-Gaussian greatly reduces the size of 3D Gaussian over 40x, and surpasses existing methods in rate-distortion performance.

Accurately and efficiently modeling dynamic scenes and motions is considered so challenging a task due to temporal dynamics and motion complexity. To address these challenges, we propose DynMF, a compact and efficient representation that decomposes a dynamic scene into a few neural trajectories. We argue that the per-point motions of a dynamic scene can be decomposed into a small set of explicit or learned trajectories. Our carefully designed neural framework consisting of a tiny set of learned basis queried only in time allows for rendering speed similar to 3D Gaussian Splatting, surpassing 120 FPS, while at the same time, requiring only double the storage compared to static scenes. Our neural representation adequately constrains the inherently underconstrained motion field of a dynamic scene leading to effective and fast optimization. This is done by biding each point to motion coefficients that enforce the per-point sharing of basis trajectories. By carefully applying a sparsity loss to the motion coefficients, we are able to disentangle the motions that comprise the scene, independently control them, and generate novel motion combinations that have never been seen before. We can reach state-of-the-art render quality within just 5 minutes of training and in less than half an hour, …

We introduce a new hair modeling method that uses a dual representation of classical hair strands and 3D Gaussians to produce accurate and realistic strand-based reconstructions from multi-view data. In contrast to recent approaches that leverage unstructured Gaussians to model human avatars, our method enforces a reconstruction of the hair in the form of 3D polylines, or strands. This fundamental difference allows us to use the resulting hairstyles out-of-the-box in modern computer graphics engines for editing, rendering, and simulation. To reconstruct strand-based hair from images, we introduce a new 3D line lifting method that utilizes unstructured Gaussians to represent the hairstyle's 3D surface. We use this intermediate reconstruction to generate multi-view geometric ground truth data to supervise the fitting of the hair strands. The hairstyle itself is represented in the form of the so-called strand-aligned 3D Gaussians. This representation allows us to combine strand-based hair priors, which are essential for realistic modeling of the inner structure of hairstyles, with the differentiable rendering capabilities of 3D Gaussian Splatting. We evaluate our method on synthetic and real hairstyles and demonstrate state-of-the-art performance in the task of strand-based hair reconstruction.

As 3D Gaussian Splatting (3DGS) provides fast and high-quality novel view synthesis, it is a natural extension to deform a canonical 3DGS to multiple frames. However, we find that previous works fail to accurately reconstruct dynamic scenes, especially 1) static parts moving along nearby dynamic parts, and 2) some motions are blurry. We attribute the failure to the wrong design of the deformation field which is built as a coordinate-based function despite 3DGS is a mixture of multiple fields centered at the Gaussians. Furthermore, the previous methods consider only single-resolution temporal embeddings. To this end, we define the deformation as a function of per-Gaussian embeddings \textit{and} temporal embeddings. Moreover, we decompose deformation as coarse and fine deformations to model slow and fast movements, respectively. Last but not least, we introduce a training strategy for faster convergence and higher quality. Code will be available online.

Synthesizing multi-view 3D from one single image is a significant but challenging task. Zero-1-to-3 methods have achieved great success by lifting a 2D latent diffusion model to the 3D scope. The target-view image is generated with a single-view source image and the camera pose as condition information. However, due to the high sparsity of the single input image, Zero-1-to-3 tends to produce geometry and appearance inconsistency across views, especially for complex objects. To tackle this issue, we propose to supply more condition information for the generation model but in a self-prompt way. A cascade framework is constructed with two Zero-1-to-3 models, named Cascade-Zero123, which progressively extract 3D information from the source image. Specifically, several nearby views are first generated by the first model and then fed into the second-stage model along with the source image as generation conditions. With amplified self-prompted condition images, our Cascade-Zero123 generates more consistent novel-view images than Zero-1-to-3. The promotion is significant for various complex and challenging scenes, involving insects, humans, transparent objects, and stacked multiple objects, etc.

Recent advances in 2D/3D generative models enable the generation of dynamic 3D objects from a single-view video. Existing approaches utilize score distillation sampling to form the dynamic scene as dynamic NeRF or dense 3D Gaussians. However, these methods struggle to strike a balance among reference view alignment, spatio-temporal consistency, and motion fidelity under single-view conditions due to the implicit nature of NeRF or the intricate dense Gaussian motion prediction. To address these issues, this paper proposes an efficient, sparse-controlled video-to-4D framework named SC4D, that decouples motion and appearance to achieve superior video-to-4D generation. Moreover, we introduce Adaptive Gaussian (AG) initialization and Gaussian Alignment (GA) loss to mitigate shape degeneration issue, ensuring the fidelity of the learned motion and shape. Comprehensive experimental results demonstrate that our method surpasses existing methods in both quality and efficiency. In addition, facilitated by the disentangled modeling of motion and appearance of SC4D, we devise a novel application that seamlessly transfers the learned motion onto a diverse array of 4D entities according to textual descriptions.

This paper presents a neural architecture MVDiffHD for 3D object reconstruction that synthesizes dense and high-resolution views of an object given one or a few images without camera poses. MVDiffHD achieves superior flexibility and scalability with two surprisingly simple ideas: 1) A pose-free architecture'' where standard self-attention among 2D latent features learns 3D consistency across an arbitrary number of conditional and generation views without explicitly using camera pose information; and 2) Aview dropout strategy'' that discards a substantial number of output views during training, which reduces the training-time memory footprint and enables dense and high-resolution view synthesis at test time. We use the Objaverse for training and the Google Scanned Objects for evaluation with standard novel view synthesis and 3D reconstruction metrics, where MVDiffHD significantly outperforms the current state of the arts. We also demonstrate a text-to-3D application example by combining MVDiffHD with a text-to-image generative model.

The increasing demand for virtual reality applications has highlighted the significance of crafting immersive 3D assets. We present a text-to-3D 360 degree scene generation pipeline that facilitates the creation of comprehensive 360 degree scenes for in-the-wild environments in a matter of minutes. Our approach utilizes the generative power of a 2D diffusion model and prompt self-refinement to create a high-quality and globally coherent panoramic image. This image acts as a preliminary "flat" (2D) scene representation. Subsequently, it is lifted into 3D Gaussians, employing splatting techniques to enable real-time exploration. To produce consistent 3D geometry, our pipeline constructs a spatially coherent structure by aligning the 2D monocular depth into a globally optimized point cloud. This point cloud serves as the initial state for the centroids of 3D Gaussians. In order to address invisible issues inherent in single-view inputs, we impose semantic and geometric constraints on both synthesized and input camera views as regularizations. These guide the optimization of Gaussians, aiding in the reconstruction of unseen regions. In summary, our method offers a globally consistent 3D scene within a 360 degree perspective, providing an enhanced immersive experience over existing techniques. We will release the code.

Feed-forward 3D generative models like the Large Reconstruction Model (LRM) have demonstrated exceptional generation speed. However, the transformer-based methods do not leverage the geometric priors of the triplane component in their architecture, often leading to sub-optimal quality given the limited size of 3D data and slow training. In this work, we present the Convolutional Reconstruction Model (CRM), a high-fidelity feed-forward single image-to-3D generative model. Recognizing the limitations posed by sparse 3D data, we highlight the necessity of integrating geometric priors into network design. CRM builds on the key observation that the visualization of triplane exhibits spatial correspondence of six orthographic images. First, it generates six orthographic view images from a single input image, then feeds these images into a convolutional U-Net, leveraging its strong pixel-level alignment capabilities and significant bandwidth to create a high-resolution triplane. CRM further employs Flexicubes as geometric representation, facilitating direct end-to-end optimization on textured meshes. Overall, our model delivers a high-fidelity textured mesh from an image in just 10 seconds, without any test-time optimization.

3D reconstruction from a sketch offers an efficient means of boosting the productivity of 3D modeling. However, such a task remains largely under-explored due to the difficulties caused by the inherent abstractive representation and diversity of sketches. In this paper, we introduce a novel deep neural network model, Sketch2Vox, for 3D reconstruction from a single monocular sketch. Taking a sketch as input, the proposed model first converts it into two different representations, i.e., a binary image and a 2D point cloud. Second, we extract semantic features from them using two newly-developed processing modules, including the SktConv module designed for hierarchical abstract features learning from the binary image and the SktMPFM designed for local and global context feature extraction from the 2D point cloud. Prior to feeding features into the 3D-decoder-refiner module for fine-grained reconstruction, the resultant image-based and point-based feature maps are fused together according to their internal correlation using the proposed cross-modal fusion attention module. Finally, we use an optimization module to refine the details of the generated 3D model. To evaluate the efficiency of our method, we collect a large dataset consisting of more than 12,000 Sketch-Voxel pairs and compare the proposed Sketch2Vox against several state-of-the-art methods. The …

We present Lagrangian Hashing, a representation for neural fields combining the characteristics of fast training NeRF methods that rely on Eulerian grids (i.e.~InstantNGP), with those that employ points equipped with features as a way to represent information (e.g. 3D Gaussian Splatting or PointNeRF). We achieve this by incorporating a point-based representation into the high-resolution layers of the hierarchical hash tables of an InstantNGP representation. As our points are equipped with a field of influence, our representation can be interpreted as a mixture of Gaussians stored within the hash table. We propose a loss that encourages the movement of our Gaussians towards regions that require more representation budget to be sufficiently well represented. Our main finding is that our representation allows the reconstruction of signals using a more compact representation without compromising quality.

We propose GaussCtrl, a text-driven method to edit a 3D scene reconstructed by the 3D Gaussian Splatting (3DGS). Our method first renders a collection of images by using the 3DGS and edits them by using a pre-trained 2D diffusion model (ControlNet) based on the input prompt, which is then used to optimise the 3D model. Our key contribution is multi-view consistent editing, which enables editing all images together instead of iteratively editing one image while updating the 3D model as in previous works. It leads to faster editing as well as higher visual quality. This is achieved by the two terms: (a) depth-conditioned editing that enforces geometric consistency across multi-view images by leveraging naturally consistent depth maps. (b) attention-based latent code alignment that unifies the appearance of edited images by conditioning their editing to several reference views through self and cross-view attention between images' latent representations. Experiments demonstrate that our method achieves faster editing and better visual results than previous state-of-the-art methods.

Recent work on image content manipulation based on vision-language pre-training models has been effectively extended to text-driven 3D scene editing. However, existing schemes for 3D scene editing still exhibit certain shortcomings, hindering their further interactive design. Such schemes typically adhere to fixed input patterns, limiting users' flexibility in text input. Moreover, their editing capabilities are constrained by a single or a few 2D visual models and require intricate pipeline design to integrate these models into 3D reconstruction processes. To address the aforementioned issues, we propose a dialogue-based 3D scene editing approach, termed CE3D, which is centered around a large language model that allows for arbitrary textual input from users and interprets their intentions, subsequently facilitating the autonomous invocation of the corresponding visual expert models. Furthermore, we design a scheme utilizing Hash-Atlas to represent 3D scene views, which transfers the editing of 3D scenes onto 2D atlas images. This design achieves complete decoupling between the 2D editing and 3D reconstruction processes, enabling CE3D to flexibly integrate a wide range of existing 2D or 3D visual models without necessitating intricate fusion designs. Experimental results demonstrate that CE3D effectively integrates multiple visual models to achieve diverse editing visual effects, possessing strong scene comprehension …

Probabilistic denoising diffusion models (DDMs) have set a new standard for 2D image generation. Extending DDMs for 3D content creation is an active field of research. Here, we propose TetraDiffusion, a diffusion model that operates on a tetrahedral partitioning of 3D space to enable efficient, high-resolution 3D shape generation. Our model introduces operators for convolution and transpose convolution that act directly on the tetrahedral partition, and seamlessly includes additional attributes like color. Our design generates mesh geometry much more efficiently: Compared to existing mesh diffusion techniques, TetraDiffusion is up to 200x faster. At the same time, it reduces memory consumption and can operate at substantially higher resolution than existing mesh generators. Using only standard consumer hardware, it sets a new standard in terms of spatial detail and outperforms other mesh generators across a range of quality metrics.

Given a 3D mesh, we aim to synthesize 3D textures that correspond to arbitrary textual descriptions. Current methods for generating and assembling textures from sampled views often result in prominent seams or excessive smoothing. To tackle these issues, we present TexGen, a novel multi-view sampling and resampling framework for texture generation leveraging a pre-trained text-to-image diffusion model. For view consistent sampling, first of all we maintain a texture map in RGB space that is parameterized by the denoising step and updated after each sampling step of the diffusion model to progressively reduce the view discrepancy. An attention-guided multi-view sampling strategy is exploited to broadcast the appearance information across views. To preserve texture details, we develop a noise resampling technique that aids in the estimation of noise, generating inputs for subsequent denoising steps, as directed by the text prompt and current texture map. Through an extensive amount of qualitative and quantitative evaluations, we demonstrate that our proposed method produces significantly better texture quality for diverse 3D objects with a high degree of view consistency and rich appearance details, outperforming current state-of-the-art methods. Furthermore, our proposed texture generation technique can also be applied to texture editing while preserving the original identity. More …

In this paper, we propose a unified framework aimed at enhancing the diffusion priors for 3D generation tasks. Despite the critical importance of these tasks, existing methodologies often struggle to generate high-caliber results. We begin by examining the inherent limitations in previous diffusion priors. We identify a divergence between the diffusion priors and the training procedures of diffusion models that substantially impairs the quality of 3D generation. To address this issue, we propose a novel, unified framework that iteratively optimizes both the 3D model and the diffusion prior. Leveraging the different learnable parameters of the diffusion prior, our approach offers multiple configurations, affording various trade-offs between performance and implementation complexity. Notably, our experimental results demonstrate that our method markedly surpasses existing techniques, establishing new state-of-the-art in the realm of text-to-3D generation. Additionally, our framework yields insightful contributions to the understanding of recent score distillation methods, such as the VSD loss and CSD loss.

Recent text-to-3D generation approaches produce impressive 3D results but require time-consuming optimization that can take up to an hour per prompt. Amortized methods like ATT3D optimize multiple prompts simultaneously to improve efficiency, enabling fast text-to-3D synthesis. However, ATT3D cannot capture high-frequency geometry and texture details and struggles to scale to large prompt sets, so it generalizes poorly. We introduce Latte3D, addressing these limitations to achieve fast, high-quality generation on a significantly larger prompt set. Key to our method is 1) building a scalable architecture for amortized learning and 2) leveraging 3D data during optimization through 3D-aware diffusion priors, shape regularization, and model initialization to achieve robustness to diverse and complex training prompts. Latte3D amortizes both neural field generation and textured surface generation to produce highly detailed textured meshes in a single forward pass. Latte3D generates 3D objects in 400ms, and can be further enhanced with fast test-time optimization.

Recent years have witnessed the strong power of 3D generation models, which offer a new level of creative flexibility by allowing users to guide the 3D content generation process through a single image or natural language. However, it remains challenging for existing 3D generation methods to create subject-driven 3D content across diverse prompts. In this paper, we introduce a novel 3D customization method, dubbed Make-Your-3D that can personalize high-fidelity and consistent 3D content from only a single image of a subject with text description within 5 minutes. Our key insight is to harmonize the distributions of a multi-view diffusion model and an identity-specific 2D generative model, aligning them with the distribution of the desired 3D subject. Specifically, we design a co-evolution framework to reduce the variance of distributions, where each model undergoes a process of learning from the other through identity-aware optimization and subject-prior optimization, respectively. Extensive experiments demonstrate that our method can produce high-quality, consistent, and subject-specific 3D content with text-driven modifications that are unseen in subject image.

Despite significant progress in generative image synthesis, and full-body generation in particular, state-of-the-art methods are either context-independent, overly reliant to text prompts, or bound to the specific, curated, training datasets, such as fashion images with monotonous backgrounds. Here, our goal is to generate people wearing clothing that is semantically appropriate for a given scene. To this end, we present ESP, a novel method for context-aware full-body generation, that enables photo-realistic inpainting of people into existing “in-the-wild” photographs. ESP is conditioned on a 2D pose and contextual cues that are extracted from the photograph of the scene and integrated into the generation process, where the clothing is modeled explicitly with human parsing masks (HPM). Generated HPMs are then used as tight guiding masks for inpainting and no changes are made to the original background. Our models are trained on a dataset containing a set of in-the-wild photographs of people covering a wide range of different environments. The method is analyzed quantitatively and qualitatively, and we show that ESP outperforms the state-of-the-art on the task of contextual full-body generation. Code will be public for research.

This study discusses the critical issues of Virtual Try-On in contemporary e-commerce and the prospective metaverse, emphasizing the challenges of preserving intricate texture details and distinctive features of the target person and the clothes in various scenarios, such as clothing texture and identity characteristics like tattoos or accessories. In addition to the fidelity of the synthesized images, the efficiency of the synthesis process presents a significant hurdle. Various existing approaches are explored, highlighting the limitations and unresolved aspects, e.g., identity information omission, uncontrollable artifacts, and low synthesis speed. It then proposes a novel diffusion-based solution that addresses garment texture preservation and user identity retention during virtual try-on. The proposed network comprises two primary modules - a warping module aligning clothing with individual features and a try-on module refining the attire and generating missing parts integrated with a mask-aware post-processing technique ensuring the integrity of the individual's identity. It demonstrates impressive results, surpassing the state-of-the-art in speed by nearly 20 times during inference, with superior fidelity in qualitative assessments. Quantitative evaluations confirm comparable performance with the recent SOTA method on the VITON-HD and Dresscode datasets.

In the realm of 3D computer vision, parametric models have emerged as a ground-breaking methodology for the creation of realistic and expressive 3D avatars. Traditionally, they rely on Principal Component Analysis (PCA), given its ability to decompose data to an orthonormal space that maximally captures shape variations. However, due to the orthogonality constraints and the global nature of PCA's decomposition, these models struggle to perform localized and disentangled editing of 3D shapes, which severely affects their use in applications requiring fine control such as face sculpting. In this paper, we leverage diffusion models to enable diverse and fully localized edits on 3D meshes, while completely preserving the un-edited regions. We propose an effective diffusion masking training strategy that, by design, facilitates localized manipulation of any shape region, without being limited to predefined regions or to sparse sets of predefined control vertices. Following our framework, a user can explicitly set their manipulation region of choice and define an arbitrary set of vertices as handles to edit a 3D mesh. Compared to the current state-of-the-art our method leads to more interpretable shape manipulations than methods relying on latent code state, greater localization and generation diversity while offering faster inference than optimization based …

This paper presents an approach to decomposing animated graphics into sprites, a set of basic elements or layers. Our approach builds on the optimization of sprite parameters to fit the raster video. For efficiency, we assume static textures for sprites to reduce the search space while preventing artifacts using a texture prior model. To further speed up the optimization, we introduce the initialization of the sprite parameters utilizing a pre-trained video object segmentation model and user input of single frame annotations. For our study, we construct the Crello Animation dataset from an online design service and define quantitative metrics to measure the quality of the extracted sprites. Experiments show that our method significantly outperforms baselines for similar decomposition tasks in terms of the quality/efficiency tradeoff.

A central problem in evolutionary biology is to explore the genetic basis of evolutionary changes in the traits of organisms, such as fin structures in fish or beak colors in birds. With the growing availability of large-scale image repositories in biology and recent advances in generative modeling, there is an opportunity to study changes in evolutionary traits of species automatically from images. We introduce a novel Hierarchical Embedding (HIER-Embed) strategy to encode the evolutionary information of a species as a composition of encodings learned at every internal node in the phylogenetic tree. We use HIER-Embeddings to condition latent diffusion models to generate synthetic images of species. Further, we introduce two novel types of perturbation operations: trait masking and trait swapping, similar in spirit to gene knockout experiments, that enable us to analyze novel changes in evolutionary traits acquired at different levels of phylogeny.

In the era of content creation revolution propelled by advancements in generative models, the field of web design remains unexplored despite its critical role in modern digital communication. The web design process is complex and often time-consuming, especially for those with limited expertise. In this paper, we introduce Web Rendering Parameters Generation (WebRPG), a new task that aims at automating the generation for visual presentation of web pages based on their HTML code. WebRPG would contribute to a faster web development workflow. Since there is no existing benchmark available, we develop a new dataset for WebRPG through an automated pipeline. Moreover, we present baseline models, utilizing VAE to manage numerous elements and rendering parameters, along with custom HTML embedding for capturing essential semantic and hierarchical information from HTML. Extensive experiments, including customized quantitative evaluations for this specific task, are conducted to evaluate the quality of the generated results.

In this paper, we introduce a novel generative model, Diffusion Layout Transformers without Autoencoder (Dolfin), which significantly improves the modeling capability with reduced complexity compared to existing methods. Dolfin employs a Transformer-based diffusion process to model layout generation. In addition to an efficient bi-directional (non-causal joint) sequence representation, we further propose an autoregressive diffusion model (Dolfin-AR) that is especially adept at capturing rich semantic correlations, such as alignment, size, overlap, and neighborhood, between layout items/elements. When evaluated against standard generative layout benchmarks, Dolfin notably improves performance across various metrics, enhancing transparency and interoperability in the process. Moreover, Dolfin's applications extend beyond layout generation, making it suitable for modeling generative geometric structures, such as line segments. Our experiments present both qualitative and quantitative results to demonstrate the advantages of Dolfin.

Diverse and realistic floor plan data are essential for the development of useful computer-aided methods in architectural design. Today's large-scale floor plan datasets predominantly feature simple floor plan layouts, typically representing single-apartment dwellings only. To compensate for the mismatch between current datasets and the real world, we develop \textbf{Modified Swiss Dwellings} (MSD) -- the first large-scale floor plan dataset that contains a significant share of layouts of multi-apartment dwellings. MSD features over 5.3K floor plans of medium- to large-scale building complexes, covering over 18.9K distinct apartments. We validate that existing approaches for floor plan generation, while effective in simpler scenarios, cannot yet seamlessly address the challenges posed by MSD. Our benchmark calls for new research in floor plan machine understanding. Code and data are open.

Accurate completion and denoising of roof height maps are crucial to reconstructing high-quality 3D buildings. Repairing sparse points can enhance low-cost sensor use and reduce UAV flight overlap. RoofDiffusion is a new end-to-end self-supervised diffusion technique for robustly completing, in particular difficult, roof height maps. RoofDiffusion leverages widely-available curated footprints and can so handle up to 99% point sparsity and 80% roof area occlusion (regional incompleteness). A variant, RoofDiffusionNF, simultaneously predicts building footprints and heights. Both quantitatively outperform state-of-the-art unguided depth completion and representative inpainting methods for Digital Elevation Models (DEM), on both a roof-specific benchmark and the BuildingNet dataset. Qualitative assessments show the effectiveness of RoofDiffusion for datasets with real-world scans including AHN3, Dales3D, and USGS 3DEP LiDAR. Tested with the leading City3D algorithm, preprocessing height maps with RoofDiffusion noticeably improves 3D building reconstruction. RoofDiffusion is complemented by a new dataset of 13k complex roof geometries, focusing on long-tail issues in remote sensing; a novel simulation of tree occlusion; and a wide variety of large-area roof cut-outs for data augmentation and benchmarking. Code and dataset will be available on GitHub.

Neural signed distance functions (SDFs) have shown powerful ability in fitting the shape geometry. However, inferring continuous signed distance fields from discrete unoriented point clouds still remains a challenge. The neural network typically fits the shape with a rough surface and omits fine-grained geometric details such as shape edges and corners. In this paper, we propose a novel non-linear implicit filter to smooth the implicit field while preserving high-frequency geometry details. Our novelty lies in that we can filter the surface (zero level set) by the neighbor input points with gradients of the signed distance field. By moving the input raw point clouds along the gradient, our proposed implicit filtering can be extended to non-zero level sets to keep the promise consistency between different level sets, which consequently results in a better regularization of the zero level set. We conduct comprehensive experiments in surface reconstruction from objects and complex scene point clouds, the numerical and visual comparisons demonstrate our improvements over the state-of-the-art methods under the widely used benchmarks.

Point cloud frame interpolation is a challenging task that involves accurate scene flow estimation across frames and maintaining the geometry structure. Prevailing techniques often rely on pre-trained motion estimators or intensive testing-time optimization, resulting in compromised interpolation accuracy or prolonged inference. This work presents FastPCI that introduces Pyramid Convolution-Transformer architecture for point cloud frame interpolation. Our hybrid Convolution-Transformer improves the local and long-range feature learning, while the pyramid network offers multilevel features and reduces the computation. In addition, FastPCI proposes a unique Dual-Direction Motion-Structure block for more accurate scene flow estimation. Our design is motivated by two facts: (1) accurate scene flow preserves 3D structure, and (2) point cloud at the previous timestep should be reconstructable using reverse motion from future timestep. Extensive experiments show that FastPCI significantly outperforms the state-of-the-art PointINet and NeuralPCI with notable gains (26.6% and 18.3% reduction in Chamfer Distance in KITTI), while being more than 10x and 600x faster, respectively.

Point clouds are commonly used in various practical applications such as autonomous driving and the manufacturing industry. However, these point clouds often suffer from incompleteness due to limited perspectives, scanner resolution and occlusion. Therefore the prediction of missing parts performs a crucial task. In this paper, we propose a novel method for point cloud completion. We utilize a spherical template to guide the generation of the coarse complete template and generate the dynamic query tokens through a correspondence pooling (Corres-Pooling) query generator. Specifically, we first generate the coarse complete template by embedding a Gaussian spherical template into the partial input and transforming the template to best match the input. Then we use the Corres-Pooling query generator to refine the coarse template and generate dynamic query tokens which could be used to predict the complete point proxies. Finally, we generate the complete point cloud with a FoldingNet following the coarse-to-fine paradigm, according to the fine template and the predicted point proxies. Experimental results demonstrate that our T-CorresNet outperforms the state-of-the-art methods on several benchmarks. We will release our code when this paper is accepted.

Recently, detection transformers (DETRs) have gradually taken a dominant position in 2D detection thanks to their elegant framework. However, DETR-based detectors for 3D point clouds are still difficult to achieve satisfactory performance. We argue that the main challenges are twofold: 1) How to obtain the appropriate object queries is challenging due to the high sparsity and uneven distribution of point clouds; 2) How to implement an effective query interaction by exploiting the rich geometric structure of point clouds is not fully explored. To this end, we propose a Simple and EffEctive 3D DETR method (SEED) for detecting 3D objects from point clouds, which involves a dual query selection (DQS) module and a deformable grid attention (DGA) module. More concretely, to obtain appropriate queries, DQS first ensures a high recall to retain a large number of queries by the predicted confidence scores and then further picks out high-quality queries according to the estimated quality scores. DGA uniformly divides each reference box into grids as the reference points and then utilizes the predicted offsets to achieve a flexible receptive field, allowing the network to focus on relevant regions and capture more informative features. Extensive ablation studies on DQS and DGA demonstrate its …

Point cloud completion aims to reconstruct the geometry of partial point clouds captured by various sensors. Traditionally, training a point cloud model is carried out on synthetic datasets, which have limited categories and deviate significantly from real-world scenarios. This disparity often leads existing methods to struggle with unfamiliar categories and severe incompleteness in real-world situations. In this paper, we propose \textbf{PrototypeCompletion}, a novel prototype-based approach for point cloud completion. It begins by generating rough prototypes and subsequently augments them with additional geometry details for the final prediction. With just a few hundred pairs of partial-complete point cloud data, our approach effectively handles the point clouds from diverse scenarios in real-world situations, including indoor ScanNet and outdoor KITTI. Additionally, we propose a new metric and test benchmark based on ScanNet200 and KITTI to evaluate the model's performance in real-world scenarios, aiming to promote future research. Experimental results demonstrate that our method outperforms state-of-the-art methods on existing PCN benchmark and excels in various real-world situations with different object categories and sensors. The code will be made available.

3D point clouds captured from real-world sensors frequently encompass noisy points due to various obstacles, such as occlusion, limited resolution, and variations in scale. This poses challenges when deploying pre-trained point cloud recognition models trained on clean point clouds, leading to significant performance degradation. While test-time adaptation (TTA) strategies have shown promising results in addressing this issue in the 2D domain, its application to 3D point clouds remains under-explored. Among TTA methods, an input adaptation approach, which directly converts test instances to the source domain using a pre-trained diffusion model, has been proposed in the 2D domain. Despite its robust TTA performance in practical situations, naively adopting this into the 3D domain may be suboptimal due to the neglect of inherent properties of point clouds, and its prohibitive computational cost. Motivated by such limitations, we propose CloudFixer, a test-time input adaptation method tailored for 3D point clouds, employing pre-trained diffusion model. Specifically, CloudFixer optimizes geometric transformation parameters with carefully designed objectives that leverage the geometric properties of point clouds. We also substantially improve computational efficiency by avoiding backpropagation through the diffusion model or extensive generation process. Furthermore, we propose an online model adaptation strategy by aligning the original model prediction …

It is challenging to reconstruct 3D point clouds in unseen classes from single 2D images. Instead of object-centered coordinate system, current methods generalized global priors learned in seen classes to reconstruct 3D shapes from unseen classes in viewer-centered coordinate system. However, the reconstruction accuracy and interpretability are still eager to get improved. To resolve this issue, we introduce to learn local pattern modularization for reconstructing 3D shapes in unseen classes, which achieves both good generalization ability and high reconstruction accuracy. Our insight is to learn a local prior which is class-agnostic and easy to generalize in object-centered coordinate system. Specifically, the local prior is learned via a process of learning and customizing local pattern modularization in seen classes. During this process, we first learn a set of patterns in local regions, which is the basis in the object-centered coordinate system to represent an arbitrary region on shapes across different classes. Then, we modularize each region on an initially reconstructed shape using the learned local patterns. Based on that, we customize the local pattern modularization using the input image by refining the reconstruction with more details. Our method enables to reconstruct high fidelity point clouds from unseen classes in object-centered coordinate …

In the realm of LiDAR-based perception, significant strides have been made, yet domain generalization remains a substantial challenge. The performance often deteriorates when models are applied to unfamiliar datasets with different LiDAR sensors or deployed in new environments, primarily due to variations in point cloud density distributions. To tackle this challenge, we propose a Density Discriminative Feature Embedding (DDFE) module, capitalizing on the observation that a single source LiDAR point cloud encompasses a spectrum of densities. The DDFE module is meticulously designed to extract density-specific features within a single source domain, facilitating the recognition of objects sharing similar density characteristics across different LiDAR sensors. In addition, we introduce a simple yet effective density augmentation technique aimed at expanding the spectrum of density in source data, thereby enhancing the capabilities of the DDFE. Our DDFE stands out as a versatile and lightweight domain generalization module. It can be seamlessly integrated into various 3D backbone networks, where it has demonstrated superior performance over current state-of-the-art domain generalization methods. We will make the source code publicly available to promote collaborative progress in the field.

Recent advancements in multi-modal pre-training for 3D point clouds have demonstrated promising results by aligning multi-modal features across 3D shapes, corresponding 2D images, and language descriptions. However, this straightforward alignment often overlooks the intricate structural relationships among the samples, potentially limiting the full capabilities of multi-modal learning. To address this issue, we introduce Multi-modal Relation Distillation (MRD), a tri-modal pretraining framework designed to effectively distill state-of-the-art large multi-modal models into 3D backbones. MRD focuses on distilling both the intra-relations within each modality and the cross-relations between different modalities, aiming to produce more discriminative 3D shape representations. Notably, MRD achieves significant improvements in downstream zero-shot classification tasks and cross-modality retrieval tasks, delivering state-of-the-art performance.

Neural fields excel in computer vision and robotics due to their ability to understand the 3D visual world such as inferring semantics, geometry, and dynamics. Given the capabilities of neural fields in densely representing a 3D scene from 2D images, we ask the question: Can we scale their self-supervised pretraining, specifically using masked autoencoders, to generate effective 3D representations from posed RGB images. Owing to the astounding success of extending transformers to novel data modalities, we employ standard 3D Vision Transformers to suit the unique formulation of NeRFs. We leverage NeRF's volumetric grid as a dense input to the transformer, contrasting it with other 3D representations such as pointclouds where the information density can be uneven, and the representation is irregular. Due to the difficulty of applying masked autoencoders to an implicit representation, such as NeRF, we opt for extracting an explicit representation that canonicalizes scenes across domains by employing the camera trajectory for sampling. Our goal is made possible by masking random patches from NeRF's radiance and density grid and employing a standard 3D Swin Transformer to reconstruct the masked patches. In doing so, the model can learn the semantic and spatial structure of complete scenes. We pretrain this …

Single-photon cameras present a promising avenue for high-resolution 3D imaging. They have ultra-high sensitivity---down to individual photons---and can record photon arrival times with extremely high (sub-nanosecond) resolution. Single-photon 3D cameras estimate the round-trip time of a laser pulse by forming equi-width (EW) histograms of detected photon timestamps. Acquiring and transferring such EW histograms requires high bandwidth and in-pixel memory, making SPCs less attractive for 3D-perception applications in resource-constrained settings such as mobile devices and AR/VR headsets. In this work we propose a new 3D sensing technique based on equi-depth (ED) histograms. ED histograms compress timestamp data more efficiently than EW histograms, reducing the bandwidth requirement. Moreover, to reduce the in-pixel memory requirement, we propose a lightweight algorithm to estimate ED histograms in an online fashion without explicitly storing the photon timestamps. This algorithm is amenable to future in-pixel implementations. We propose algorithms that process ED histograms to perform 3D computer-vision tasks of estimating scene distance maps and performing visual odometry under challenging conditions such as high ambient light. Our work paves the way towards lower bandwidth and reduced in-pixel memory requirements for SPCs, making them attractive for resource-constrained 3D vision applications.

Initialization-free bundle adjustment (BA) remains largely uncharted. While Levenberg-Marquardt algorithm is the golden method to solve the BA problem, it generally relies on a good initialization. In contrast, the under-explored Variable Projection algorithm (VarPro) exhibits a wide convergence basin even without initialization. Coupled with object space error formulation, recent works have shown its ability to solve (small-scale) initialization-free bundle adjustment problem. We introduce Power Variable Projection (PoVar), extending a recent inverse expansion method based on power series. Importantly, we link the power series expansion to Riemannian manifold optimization. This projective framework is crucial to solve large-scale bundle adjustment problem without initialization. Using the real-world BAL dataset, we experimentally demonstrate that our solver achieves state-of-the-art results in terms of speed and accuracy. In particular, our work is the first, to our knowledge, that addresses the scalability of BA without initialization and opens new venues for initialization-free Structure-from-Motion.

Unsupervised 3D keypoints estimation from Point Cloud Data (PCD) is a complex task, even more challenging when an object shape is deforming. As keypoints should be semantically and geometrically consistent across all the 3D frames -- each keypoint should be anchored to a specific part of the deforming shape irrespective of intrinsic and extrinsic motion. This paper presents, "SelfGeo", a self-supervised method that computes persistent 3D keypoints of non-rigid objects from arbitrary PCDs without the need of human annotations. The gist of SelfGeo is to estimate keypoints between frames that respect invariant properties of deforming bodies. Our main contribution is to enforce that keypoints deform along with the shape while keeping constant geodesic distances among them. This principle is then propagated to the design of a set of losses which minimization let emerge repeatable keypoints in specific semantic locations of the non-rigid shape. We show experimentally that the use of geodesic has a clear advantage in challenging dynamic scenes and with different classes of deforming shapes (humans and animals). Code and data will be made available upon paper acceptance.

In this paper, we derive a linear constraint for planar motion leveraging scale- and orientation covariant features, e.g., SIFT, which is used to create a novel minimal solver for planar motion requiring only a single covariant feature. We compare the proposed method to traditional point-based solvers and solvers relying on affine correspondences in controlled synthetic environments and well-established datasets for autonomous driving. The proposed solver is integrated in a modern robust estimation framework, where it is shown to accelerate the complete estimation pipeline more than 25x, compared to state-of-the-art affine-based minimal solvers, with negligible loss in precision.

Simultaneous localization and mapping (SLAM) with implicit neural representations has received extensive attention due to the expressive representation power and the innovative paradigm of continual learning. However, deploying such a system within a dynamic environment has not been well-studied. Such challenges are intractable even for conventional algorithms since observations from different views with dynamic objects involved break the geometric and photometric consistency, whereas the consistency lays the foundation for joint optimizing the camera pose and the map parameters. In this paper, we best exploit the characteristics of continual learning and propose a novel SLAM framework for dynamic environments. While past efforts have been made to avoid catastrophic forgetting by exploiting an experience replay strategy, we view forgetting as a desirable characteristic: by adaptively controlling the replayed buffer, the ambiguity caused by moving objects can be easily alleviated through forgetting. We restrain the replay of the dynamic objects by introducing a continually-learned classifier for dynamic object identification. The iterative optimization of the neural map and the classifier notably improves the robustness of the SLAM system under a dynamic environment. Experiments on challenging datasets verify the effectiveness of the proposed framework.

Recent advances in interactive keypoint estimation methods have enhanced keypoint estimation accuracy while aiming to minimize user intervention. However, these methods still depend on user input for error correction, a significant challenge in vertebrae keypoint estimation where densely clustered or overlapping keypoints are common. We introduce a novel two-stage approach that integrates KeyBot that specifically designed to identify and correct significant errors in existing models into existing keypoint estimation frameworks. It is specifically designed to analyze current model predictions and deliver corrective feedback akin to user revision. Trained on simulated error scenarios, KeyBot effectively corrects typical errors in vertebrae keypoint estimation, thereby significantly reducing user workload. Comprehensive quantitative and qualitative evaluations on three public datasets confirm that KeyBot significantly outperforms existing methods, achieving state-of-the-art performance in interactive vertebrae keypoint estimation.

Assembling 3D objects from primitive bricks is challenging due to complex constraints and numerous possible combinations. Recent studies have demonstrated promising results on sequential LEGO brick assembly by graph modeling. However, existing approaches are class-specific and require significant computational and 3D annotation resources. In this work, we first propose a computationally efficient breadth-first search (BFS) LEGO-Tree structure to model sequential assembly actions. Based on the LEGO-Tree structure, we then design a class-agnostic tree-transformer framework to predict assembly actions from multi-view images. A major challenge is the costly acquisition of step-wise action labels. We address this by leveraging synthetic-to-real transfer learning. Specifically, our model pre-trains on synthetic data with full action label supervision. We further circumvent the requirement for real data action labels by introducing an action-to-silhouette projection for self-supervision. With no real data annotation, our model outperforms existing methods with 3D supervision by 7.8% and 11.3% in mIoU on the MNIST and ModelNet Construction datasets, respectively.

Monocular 3D reconstruction for categorical objects heavily relies on accurately perceiving each object's pose. While gradient-based optimization within a NeRF framework updates initially given poses, this paper highlights that such a scheme fails when the initial pose even moderately deviates from the true pose. Consequently, existing methods often depend on a third-party 3D object to provide an initial object pose, leading to increased complexity and generalization issues. To address these challenges, we present UPNeRF, a Unified network integrating Pose estimation and NeRF-based reconstruction, bringing us closer to real-time monocular 3D object reconstruction. UPNeRF decouples the object's dimension estimation and pose refinement to resolve the scale-depth ambiguity, and introduces an effective projected-box representation that generalizes well cross different domains. While using a dedicated pose estimator that smoothly integrates into an object-centric NeRF , UPNeRF is free from external 3D detectors. UPNeRF achieves state-of-the-art results in both reconstruction and pose estimation tasks on the nuScenes dataset. Furthermore, UPNeRF exhibits exceptional Cross-dataset generalization on the KITTI and Waymo datasets, surpassing prior methods with up to 50\% reduction in rotation and translation error.

Previous works on Human Pose and Shape Estimation (HPSE) from RGB images can be broadly categorized into two main groups: parametric and non-parametric approaches. Parametric techniques leverage a low-dimensional statistical body model for realistic results, whereas recent non-parametric methods achieve higher precision by directly regressing the 3D coordinates of the human body mesh. This work introduces a novel paradigm to address the HPSE problem, involving a low-dimensional discrete latent representation of the human mesh and framing HPSE as a classification task. Instead of predicting body model parameters or 3D vertex coordinates, we focus on predicting the proposed discrete latent representation, which can be decoded into a registered human mesh. This innovative paradigm offers two key advantages. Firstly, predicting a low-dimensional discrete representation confines our predictions to the space of anthropomorphic poses and shapes even when little training data is available. Secondly, by framing the problem as a classification task, we can harness the discriminative power inherent in neural networks. The proposed model, VQ-HPS, predicts the discrete latent representation of the mesh. The experimental results demonstrate that VQ-HPS outperforms the current state-of-the-art non-parametric approaches while yielding results as realistic as those produced by parametric methods when trained with few data. VQ-HPS …

Existing 2D-to-3D pose lifting networks suffer from poor performance in cross-dataset benchmarks. Although 2D keypoints joined by "stick-figure’’ limbs is the dominant trend, stick-figures do not preserve occlusion information that is inherent in an image, resulting in significant ambiguities that are ruled out when occlusion information is present. In addition, datasets with ground truth 3D poses are much harder to obtain in contrast to similar human annotated 2D datasets. To address these issues, we propose to replace stick figures with abstract images ---figures with opaque limbs that preserve occlusion information while implicitly encoding joint locations. We then break down the pose estimation task into two stages: (1) Generating an abstract image from a real image, and (2) garnering the pose from the abstract image. Crucially, given the GT 3D keypoints for a particular pose, we can synthesize an arbitrary number of abstract images of the same pose as seen from arbitrary cameras, even without a part map. Given a set of 3D GT keypoints, this allows training of Stage (2) on an unlimited dataset without over-training, which in turn allows us to correctly interpret poses from arbitrary viewpoints not included in the original dataset. Additionally, our unlimited training of Stage …

Traditional methods for human localization and pose estimation (HPE), which mainly rely on RGB images as an input modality, confront substantial limitations in real-world applications due to privacy concerns. In contrast, radar-based HPE methods emerge as a promising alternative, characterized by distinctive attributes such as through-wall recognition and privacy-preserving, rendering the method more conducive to practical deployments. This paper presents a Radar Tensor-based human pose (RT-Pose) dataset and an open-source benchmarking framework. RT-Pose dataset comprises 4D radar tensors, LiDAR point clouds, and RGB images, and is collected for a total of 72k frames across 240 sequences with six different complexity level actions. The 4D radar tensor provides raw spatio-temporal information, differentiating it from other radar point cloud-based datasets. We develop a semi-automatic annotation process, which uses RGB images and LiDAR point clouds to accurately label 3D human skeletons.In addition, we propose HRRadarPose, the first single-stage architecture that extracts the high-resolution representation of 4D radar tensors in 3D space to aid human keypoint estimation. HRRadarPose outperforms previous radar-based HPE work on the RT-Pose benchmark. The overall HRRadarPose performance on the RT-Pose dataset, as reflected in a mean per joint position error (MPJPE) of 9.91 cm, indicates the persistent challenges in achieving …

This study addresses the nuanced challenge of estimating head translations within the context of six-degrees-of-freedom (6DoF) head pose estimation, placing emphasis on this aspect over the more commonly studied head rotations. Identifying a gap in existing methodologies, we recognized the underutilized potential synergy between facial geometry and head translation. To bridge this gap, we propose a novel approach called the head \textbf{T}ranslation, \textbf{R}otation, and face \textbf{G}eometry network (TRG), which stands out for its explicit bidirectional interaction structure. This structure has been carefully designed to leverage the complementary relationship between face geometry and head translation, marking a significant advancement in the field of head pose estimation. Our contributions also include the development of a strategy for estimating bounding box correction parameters and a technique for aligning landmarks to image. Both of these innovations demonstrate superior performance in 6DoF head pose estimation tasks. Extensive experiments conducted on ARKitFace and BIWI datasets confirm that the proposed method outperforms current state-of-the-art techniques. The code will be released.

Predicting camera-space hand meshes from single RGB images is crucial for enabling realistic hand interactions in 3D virtual and augmented worlds. Previous works typically divided the task into two stages: given a cropped image of the hand, predict meshes in relative coordinates, followed by lifting these predictions into camera space in a separate and independent stage, often resulting in the loss of valuable contextual and scale information. To prevent the loss of these cues, we propose unifying these two stages into an end-to-end solution that addresses the 2D-3D correspondence problem. This solution enables back-propagation from camera space outputs to the rest of the network through a new differentiable global positioning module. Additionally, we introduce an image rectification step that harmonizes both the training dataset and the input image as if they were acquired with the same camera, helping to alleviate the inherent scale-depth ambiguity of the problem. We validate the effectiveness of our framework in evaluations against several baselines and state-of-the-art approaches across three public benchmarks. The code and models will be made available.

3D hand poses are an under-explored modality for action recognition. Poses are compact yet informative and can greatly benefit applications with limited compute budgets. However, poses alone offer an incomplete understanding of actions, as they cannot fully capture objects and environments with which humans interact. To efficiently model hand-object interactions, we propose HandFormer, a novel multimodal transformer. HandFormer combines 3D hand poses at a high temporal resolution for fine-grained motion modeling with sparsely sampled RGB frames for encoding scene semantics. Observing the unique characteristics of hand poses, we temporally factorize hand modeling and represent each joint by its short-term trajectories. This factorized pose representation combined with sparse RGB samples is remarkably efficient and achieves high accuracy. Unimodal HandFormer with only hand poses outperforms existing skeleton-based methods at 5x fewer FLOPs. With RGB, we achieve new state-of-the-art performance on Assembly101 and H2O with significant improvements in egocentric action recognition.

Human Pose Forecasting is a major problem in human intention comprehension that can be addressed through learning the historical poses via deep methods. However, existing methods often lack the modeling of the person's role in the event in multi-person scenes. This leads to limited performance in complicated scenes with variant interactions happening at the same time. In this paper, we introduce the Interaction-Aware Pose Forecasting Transformer (IAFormer) framework to better learn the interaction features. With the key insight that the event often involves only part of the people in the scene, we designed the Interaction Perceptron Module (IPM) to evaluate the human-to-event interaction level. With the interaction evaluation, the human-independent features are extracted with the attention mechanism for interaction-aware forecasting. In addition, an Interaction Prior Learning Module (IPLM) is presented to learn and accumulate prior knowledge of high-frequency interactions, encouraging semantic pose forecasting rather than simple trajectory pose forecasting. We conduct experiments using datasets such as CMU-Mocap, UMPM, CHI3D, Human3.6M datasets. The results demonstrate that our method significantly outperforms state-of-the-art approaches considering scenarios with varying numbers of people.

Performing language-conditioned robotic manipulation tasks in unstructured environments is highly demanded for general intelligent robots. Conventional robotic manipulation methods usually learn semantic representation of the observation for action prediction, which ignores the scene-level spatiotemporal dynamics for human goal completion. In this paper, we propose a dynamic Gaussian Splatting method named ManiGaussian for multi-task robotic manipulation, which mines scene dynamics via future scene reconstruction. Specifically, we first formulate the dynamic Gaussian Splatting framework that infers the semantics propagation in the Gaussian embedding space, where the semantic representation is leveraged to predict the optimal robot action. Then, we build a Gaussian world model to parameterize the distribution in our dynamic Gaussian Splatting framework, which provides informative supervision in the interactive environment via future scene reconstruction. We evaluate our ManiGaussian on 10 RLBench tasks with 166 variations, and the results demonstrate our framework can outperform the state-of-the-art methods by 13.1\% in average success rate.

Unsupervised depth completion methods are trained by minimizing sparse depth and image reconstruction error. Block artifacts from resampling, intensity saturation, and occlusions are amongst the many undesirable by-products of common data augmentation schemes that affect image reconstruction quality, and thus the training signal. Hence, typical augmentations on images viewed as essential to training pipelines in other vision tasks have seen limited use beyond small image intensity changes and flipping. The sparse depth modality have seen even less as intensity transformations alter the scale of the 3D scene, and geometric transformations may decimate the sparse points during resampling. We propose a method that unlocks a wide range of previously-infeasible geometric augmentations for unsupervised depth completion. This is achieved by reversing, or undo"-ing, geometric transformations to the coordinates of the output depth, warping the depth map back to the original reference frame. This enables computing the reconstruction losses using the original images and sparse depth maps, eliminating the pitfalls of naive loss computation on the augmented inputs. This simple yet effective strategy allows us to scale up augmentations to boost performance. We demonstrate our method on indoor (VOID) and outdoor (KITTI) datasets where we improve upon recent methods by an average of …

In the area of self-supervised monocular depth estimation, models that utilize rich-resource inputs, such as high-resolution and multi-frame inputs, typically achieve better performance than models that use ordinary single image input. However, these rich-resource inputs may not always be available, limiting the applicability of these methods in general scenarios. In this paper, we propose Rich-resource Prior Depth estimator (RPrDepth), which only requires single input image during the inference phase but can still produce highly accurate depth estimations comparable to rich-resource based methods. Specifically, we consider rich-resource inputs as a prior information and extract rich-resource features from them during the training phase. When estimating the depth for a single-image image, we search for similar pixels from the rich-resource features and use them as prior information to estimate the depth. Experimental results demonstrate that our model outperform other single-image model and can achieve comparable or even better performance than models with rich-resource inputs, only using low-resolution single-image input.

The ground-to-satellite image matching/retrieval was initially proposed for city-scale ground camera localization. This work addresses the problem of improving camera pose accuracy by ground-to-satellite image matching after a coarse location and orientation have been obtained, either from the city-scale retrieval or from consumer-level GPS and compass sensors. Existing learning-based methods for solving this task require accurate GPS labels of ground images for network training. However, obtaining such accurate GPS labels is difficult, often requiring an expensive Real Time Kinematics (RTK) setup and suffering from signal occlusion, multi-path signal disruptions, \etc. To alleviate this issue, this paper proposes a weakly supervised learning strategy for ground-to-satellite image registration when only noisy pose labels for ground images are available for network training. It derives positive and negative satellite images for each ground image and leverages contrastive learning to learn feature representations for ground and satellite images useful for translation estimation. We also propose a self-supervision strategy for cross-view image relative rotation estimation, which trains the network by creating pseudo query and reference image pairs. Experimental results show that our weakly supervised learning strategy achieves the best performance on cross-area evaluation compared to recent state-of-the-art methods that are reliant on accurate pose labels for …

Cross-view geo-localization serves as a viable alternative to providing geographical location information when GPS signals are unstable or unavailable by matching ground images with geo-tagged aerial image databases. While significant progress has been made on some common benchmarks like CVUSA and CVACT, there remains a lack of comprehensive consideration for robustness against real-world environmental challenges such as adverse weather or sensor noise. This deficiency poses a significant challenge for deploying this technology in safety-critical domains like autonomous driving and robot navigation. To the best of our knowledge, there is currently no specialized benchmark for evaluating the robustness of cross-view geo-localization. To comprehensively and fairly evaluate the robustness of cross-view geo-localization models in real-world scenarios, we introduce 16 common types of data corruption. By synthesizing these corruptions on public datasets, we establish two fine-grained corruption robustness benchmarks (CVUSA-C and CVACTval-C) and three comprehensive corruption robustness benchmarks (CVUSA-C-ALL, CVACTval-C-ALL, and CVACT_test-C-ALL), covering approximately 1.5 million corrupted images. Subsequently, we conduct large-scale experiments on various cross-view geo-localization models to evaluate their robustness in corrupted environments and derive novel insights. Finally, we explore two data augmentation strategies as potential solutions to enhance model robustness. Combined with the training strategies proposed, these approaches …

Crowd counting and localization have become increasingly important in computer vision due to their wide-ranging applications. While point-based strategies have been widely used in crowd counting methods, they face a significant challenge, i.e., the lack of an effective learning strategy to guide the matching process. This deficiency leads to instability in matching point proposals to target points, adversely affecting overall performance. To address this issue, we introduce an effective approach to stabilize the proposal-target matching in point-based methods. We propose Auxiliary Point Guidance (APG) to provide clear and effective guidance for proposal selection and optimization, addressing the core issue of matching uncertainty. Additionally, we develop Implicit Feature Interpolation (IFI) to enable adaptive feature extraction in diverse crowd scenarios, further enhancing the model's robustness and accuracy. Extensive experiments demonstrate the effectiveness of our approach, showing significant improvements in crowd counting and localization performance, particularly under challenging conditions. The source codes and trained models will be made publicly available.

The Bird's-Eye-View (BEV) representation is a critical factor that directly impacts the 3D object detection performance, but the traditional BEV grid representation induces quadratic computational cost as the spatial resolution grows. To address this limitation, we present a new camera-based 3D object detector with high-resolution vector representation: VectorFormer. The presented high-resolution vector representation is combined with the lower-resolution BEV representation to efficiently exploit 3D geometry from multi-camera images at a high resolution through our two novel modules: vector scattering and gathering. To this end, the learned vector representation with richer scene contexts can serve as the decoding query for final predictions. We conduct extensive experiments on the nuScenes dataset and demonstrate state-of-the-art performance in NDS and inference time. Furthermore, we investigate query-BEV-based methods incorporated with our proposed vector representation and observe a consistent performance improvement. Our source code will be made publicly available.

Integrating LiDAR and camera information into Bird's-Eye-View (BEV) representations has emerged as a crucial aspect of 3D object detection in autonomous driving. However, existing methods are susceptible to the inaccurate calibration relationship between LiDAR and the camera sensor. Such inaccuracies result in errors in depth estimation for the camera branch, ultimately causing misalignment between LiDAR and camera BEV features. In this work, we propose a robust fusion framework called GraphBEV. Addressing errors caused by inaccurate point cloud projection, we introduce a LocalAlign module that employs neighbor-aware depth features via Graph matching. Additionally, we propose a GlobalAlign module to rectify the misalignment between LiDAR and camera BEV features. Our GraphBEV framework achieves state-of-the-art performance, with an mAP of 70.1%, surpassing BEVFusion by 1.6% on the nuScnes validation set. Importantly, our GraphBEV outperforms BEVFusion by 8.3% under conditions with misalignment noise.

3D single object tracking (SOT) is an essential task in autonomous driving and robotics. However, learning robust 3D SOT trackers remains challenging due to the limited category-specific point cloud data and the inherent sparsity and incompleteness of LiDAR scans. To tackle these issues, we propose a unified 3D SOT framework that leverages 3D generative pre-training and learns robust 3D matching abilities from 2D pre-trained foundation trackers. Our framework features a consistent target-matching architecture with the widely used 2D trackers, facilitating the transfer of 2D matching knowledge. Specifically, we first propose a lightweight Target-Aware Projection (TAP) module, allowing the pre-trained 2D tracker to work well on the projected point clouds without further fine-tuning. We then propose a novel IoU-guided matching-distillation framework that utilizes the powerful 2D pre-trained trackers to guide 3D matching learning in the 3D tracker, i.e., the 3D template-to-search matching should be consistent with its corresponding 2D template-to-search matching obtained from 2D pre-trained trackers. Our designs are applied to two mainstream 3D SOT frameworks: memory-less Siamese and contextual memory-based approaches, which are respectively named SiamDisst and MemDisst. Extensive experiments show that SiamDisst and MemDisst achieve state-of-the-art performance on KITTI, Waymo Open Dataset and nuScenes benchmarks, while running at the …

Accurate 3D object detection is vital for automated driving perception. While lidar sensors are well suited for this task, they are expensive and have limitations in adverse weather conditions. 3+1D imaging radar sensors offer a cost-effective, robust alternative but face challenges due to their low resolution and high measurement noise. Existing 3+1D imaging radar datasets include both radar and lidar data, enabling cross-modal model improvements. Although lidar shall not be used during inference, it can aid the training of a radar-only object detector. We explore two strategies to transfer knowledge from the lidar to the radar domain and radar-only object detectors: 1. multi-stage training with sequential lidar point cloud thin-out, and 2. cross-modal knowledge distillation. In the multi-stage process, three thin-out methods are examined. Our results show significant performance gains of up to 4.2 percentage points in mean Average Precision with multi-stage training and up to 3.9 percentage points with knowledge distillation by initializing the student with the teacher's weights. The main benefit of these approaches is their applicability to other 3D object detection networks without altering their architecture, as we show by analyzing it on two different object detectors.

In autonomous driving, the temporal stability of 3D object detection greatly impacts the driving safety. However, the detection stability cannot be accessed by existing metrics such as mAP and MOTA, and consequently is less explored by the community. To bridge this gap, this work proposes Stability Index (SI), a new metric that can comprehensively evaluate the stability of 3D detectors in terms of confidence, box localization, extent, and heading. By benchmarking state-of-the-art object detectors on the Waymo Open Dataset, SI reveals interesting properties of object stability that have not been previously discovered by other metrics. To help models improve their stability, we further introduce a general and effective training strategy, called Prediction Consistency Learning (PCL). PCL essentially encourages the prediction consistency of the same objects under different timestamps and augmentations, leading to enhanced detection stability. Furthermore, we examine the effectiveness of PCL with the widely-used CenterPoint, and achieve a remarkable SI of 86.00 for vehicle class, surpassing the baseline by 5.48. We hope our work could serve as a reliable baseline and draw the community's attention to this crucial issue in 3D object detection. Codes will be made publicly available.

3D occupancy, an advanced perception technology for driving scenarios, represents the entire scene without distinguishing between foreground and background by quantifying the physical space into a grid map. The widely adopted projection-first deformable attention, efficient in transforming image features into 3D representations, encounters challenges in aggregating multi-view features due to sensor deployment constraints. To address this issue, we propose our learning-first view attention mechanism for effective multi-view feature aggregation. Moreover, we showcase the scalability of our view attention across diverse multi-view 3D tasks, such as map construction and 3D object detection. Leveraging the proposed view attention as well as an additional multi-frame streaming temporal attention, we introduce ViewFormer, a vision-centric transformer-based framework for spatiotemporal feature aggregation. To further explore occupancy-level flow representation, we present FlowOcc3D, a benchmark built on top of existing high-quality datasets. Qualitative and quantitative analyses on this benchmark reveal the potential to represent fine-grained dynamic scenes. Extensive experiments show that our approach significantly outperforms prior state-of-the-art methods. The codes and benchmark will be released soon.

Animals perceive the world to plan their actions and interact with other agents to accomplish complex tasks, demonstrating capabilities that are still unmatched by AI systems. To advance our understanding and reduce the gap between the capabilities of animals and AI systems, we introduce a dataset of pet egomotion imagery with diverse examples of simultaneous egomotion and multi-agent interaction. Current video datasets separately contain egomotion and interaction examples, but rarely both at the same time. In addition, EgoPet offers a radically distinct perspective from existing egocentric datasets of humans or vehicles. We define two in-domain benchmark tasks that capture animal behavior, and a third benchmark to assess the utility of EgoPet as a pretraining resource to robotic quadruped locomotion, showing that models trained from EgoPet outperform those trained from prior datasets. This work provides evidence that today's pets could be a valuable resource for training future AI systems and robotic assistants.

Generating multi-camera street-view videos is critical for augmenting autonomous driving datasets, addressing the urgent demand for extensive and varied data. Due to the limitations in diversity and challenges in handling lighting conditions, traditional rendering-based methods are increasingly being supplanted by diffusion-based methods. However, a significant challenge in diffusion-based methods is ensuring that the generated sensor data preserve both intra-world consistency and inter-sensor coherence. To address these challenges, we combine an additional explicit world volume and propose the World Volume-aware Multi-camera Driving Scene Generator (WoVoGen). This system is specifically designed to leverage 4D world volume as a foundational element for video generation. Our model operates in two distinct phases: (i) envisioning the future 4D temporal world volume based on vehicle control sequences, and (ii) generating multi-camera videos, informed by this envisioned 4D temporal world volume and sensor interconnectivity. The incorporation of the 4D world volume empowers WoVoGen not only to generate high-quality street-view videos in response to vehicle control inputs but also to facilitate scene editing tasks.

Predicting the maneuvers of surrounding vehicles is imperative for the safe navigation of autonomous vehicles. However, naturalistic driving datasets tend to be highly imbalanced, with a bias towards the "going straight" maneuver. Consequently, learning and accurately predicting turning maneuvers pose significant challenges. In this study, we propose a novel two-stage maneuver learning method that can overcome such strong biases by leveraging two heterogeneous datasets in a complementary manner. In the first training phase, we utilize an intersection-centric dataset characterized by balanced distribution of maneuver classes to learn the representations of each maneuver. Subsequently, in the second training phase, we incorporate an ego-centric driving dataset to account for various geometrical road shapes, by transferring the knowledge of geometric diversity to the maneuver prediction model. To facilitate this, we constructed an in-house intersection-centric trajectory dataset with a well-balanced maneuver distribution. By harnessing the power of heterogeneous datasets, our framework significantly improves maneuver prediction performance, particularly for minority maneuver classes such as turning maneuvers. The dataset will be made publicly available soon.

3D semantic occupancy prediction aims to obtain 3D fine-grained geometry and semantics of the surrounding scene and is an important task for the robustness of vision-centric autonomous driving. Most existing methods employ dense grids such as voxels as scene representations, which ignore the sparsity of occupancy and the diversity of object scales and thus lead to unbalanced allocation of resources. To address this, we propose an object-centric representation to describe 3D scenes with sparse 3D semantic Gaussians where each Gaussian represents a flexible region of interest and its semantic features. We aggregate information from images through the attention mechanism and iteratively refine the properties of 3D Gaussians including position, covariance, and semantics. We then propose an efficient Gaussian-to-voxel splatting method to generate 3D occupancy predictions, which only aggregates the neighboring Gaussians for a certain position. We conduct extensive experiments on the widely adopted nuScenes and KITTI-360 datasets. Experimental results demonstrate that GaussianFormer achieves comparable performance with state-of-the-art methods with only 17.8% - 24.8% of their memory consumption.

In the field of autonomous driving, online High-definition (HD) map construction is crucial for planning tasks. Recent studies have developed several high-performance HD map construction models to meet the demand. However, the point sequences generated by recent HD map construction models are jittery or jagged due to prediction bias and impact subsequent tasks. To mitigate this jitter issue, we propose the Anti-Disturbance Map construction framework (ADMap), which contains Multi-scale Perception Neck (MPN), Instance Interactive Attention (IIA), and Vector Direction Difference Loss (VDDL). By exploring the point sequence relations between and within instances in a cascading manner, our proposed ADMap effectively monitors the point sequence prediction process, and achieves state-of-the-art performance on the nuScenes and Argoverse2 datasets. Extensive results demonstrate its ability to produce stable and reliable map elements in complex and changing driving scenarios.

Online lane graph construction is a promising but challenging task in autonomous driving. Previous methods usually model the lane graph at the pixel or piece level, and recover the lane graph by pixel-wise or piece-wise connection, which breaks down the continuity of the lane and results in suboptimal performance. Human drivers focus on and drive along the continuous and complete paths instead of considering lane pieces. Autonomous vehicles also require path-specific guidance from lane graph for trajectory planning. We argue that the path, which indicates the traffic flow, is the primitive of the lane graph. Motivated by this, we propose to model the lane graph in a novel path-wise manner, which well preserves the continuity of the lane and encodes traffic information for planning. We present a path-based online lane graph construction method, termed LaneGAP, which end-to-end learns the path and recovers the lane graph via a Path2Graph algorithm. We qualitatively and quantitatively demonstrate the superior accuracy and efficiency of LaneGAP over conventional pixel-based and piece-based methods on the challenging nuScenes and Argoverse2 datasets under controllable and fair conditions. Compared to the recent state-of-the-art piece-wise method TopoNet on the OpenLane-V2 dataset, LaneGAP still outperforms by 1.6 mIoU, further validating the …

The choice of representation plays a key role in self-driving. Bird’s eye view (BEV) representations have shown remarkable performance in recent years. In this paper, we propose to learn object-centric representations in BEV to distill a complex scene into more actionable information for self-driving. We first learn to place objects into slots with a slot attention model on BEV sequences. Based on these object-centric representations, we then train a transformer to learn to drive as well as reason about the future of other vehicles. We found that object-centric slot representations outperform both scene-level and object-level approaches that use the exact attributes of objects. Slot representations naturally incorporate information about objects from their spatial and temporal context such as position, heading, and speed without explicitly providing it. Our model with slots achieves increased coverage of the provided routes and, consequently, a higher driving score, with a lower variance across multiple runs, affirming slots as a reliable alternative in object-centric approaches. Additionally, we validate our model’s performance as a world model through forecasting experiments, demonstrating its capability to accurately predict future slot representations.

The lack of generalization capability of behavior prediction models for autonomous vehicles is a crucial concern for safe motion planning. One way to address this is via self-supervised pre-training through masked trajectory prediction. However, the existing models rely on uniform random sampling of tokens, which is sub-optimal because it implies that all components of driving scenes are equally informative. In this paper, to enable more robust representation learning, we introduce a dynamic masked self-distillation approach to identify and utilize informative aspects of the scenes, particularly those corresponding to complex driving behaviors, such as overtaking. Specifically, for targeted sampling, we propose a dynamic method that prioritizes tokens, such as trajectory or lane segments, based on their informativeness. The latter is determined via an auxiliary network that estimates token distributions. Through sampler optimization, more informative tokens are rewarded and selected as visible based on the policy gradient algorithm adopted from reinforcement learning. In addition, we propose a masked self-distillation approach to transfer knowledge from fully visible to masked scene representations. The distillation process not only enriches the semantic information within the visible token set but also progressively refines the sampling process. Further, we use an integrated training regime to enhance the model's …

We present a versatile NeRF-based simulator for testing autonomous driving (AD) software systems, designed with a focus on sensor-realistic closed-loop evaluation and the creation of safety-critical scenarios. The simulator learns from sequences of real-world driving sensor data and enables reconfigurations and renderings of new, unseen scenarios. In this work, we use our simulator to test the responses of AD models to safety-critical scenarios inspired by the European New Car Assessment Programme (Euro NCAP). Our evaluation reveals that, while state-of-the-art end-to-end planners excel in nominal driving scenarios in an open-loop setting, they exhibit critical flaws when navigating our safety-critical scenarios in a closed-loop setting. This highlights the need for advancements in the safety and real-world usability of end-to-end planners. By publicly releasing our simulator and scenarios as an easy-to-run evaluation suite, we invite the research community to explore, refine, and validate their AD models in controlled, yet highly configurable and challenging sensor-realistic environments.

What will be the relationships between objects in a novel view? We strive to answer this question by investigating a new visual cognition task, termed visual relationship transformation or VRT. Unlike prior visual relationship detection task that works on visible view images, VRT aims to predict the relationships in unseen novel views from a single observed source view. Towards solving VRT, we propose an end-to-end deep approach that, given an observed view image and inter-view transformations, learns to predict the relationships in novel views. Specifically, we introduce an equivariant graph neural network to predict the relationships between objects in novel views, which is achieved by enforcing the transformation equivariance of the learned relationship representations. Simultaneously, a relationship presentness mask is learned for pruning the invisible ones, thus enabling the visible relationship prediction in novel views. To this end, VRT provides supplementary cues for accomplishing novel-view-related tasks, such as visual grounding (VG), novel view synthesis (NVS), and pedestrian intention estimation (PIE). In the experiments, adopting VRT as a plug-in module results in considerable performance improvements in VG, NVS, and PIE across all datasets.

We introduce LocoTrack, a highly accurate and efficient model designed for the task of tracking any point (TAP) across video sequences. Previous approaches in this task often rely on local 2D correlation maps to establish correspondences from a point in the query image to a local region in the target image, which often struggle with homogeneous regions or repetitive features, leading to matching ambiguities. LocoTrack overcomes this challenge with a novel approach that utilizes all-pair correspondences across regions, i.e., local 4D correlation, to establish precise correspondences, with bidirectional correspondence and matching smoothness significantly enhancing robustness against ambiguities. We also incorporate a lightweight correlation encoder to enhance computational efficiency, and a compact Transformer architecture to integrate long-term temporal information. LocoTrack achieves unmatched accuracy on all TAP-Vid benchmarks and operates at a speed almost 5x faster than the current state-of-the-art.

Event cameras are a novel type of biologically inspired vision sensor known for their high temporal resolution, high dynamic range, and low power consumption. Because of these properties, they are well-suited for processing fast motions that require rapid reactions. Although event cameras have recently shown competitive performance in unsupervised optical flow estimation, performance in detecting independently moving objects (IMOs) is lacking behind, although event-based methods would be suited for this task based on their low latency and HDR properties. Previous approaches to event-based IMO segmentation have been heavily dependent on labeled data. However, biological vision systems have developed the ability to avoid moving objects through daily tasks without being given explicit labels. In this work, we propose the first event framework that generates IMO pseudo-labels using geometric constraints. Due to its unsupervised nature, our method can handle an arbitrary number of not predetermined objects and is easily scalable to datasets where expensive IMO labels are not readily available. We evaluate our approach on the EVIMO dataset and show that it performs competitively with supervised methods, both quantitatively and qualitatively.

Traditional frame-based cameras have achieved impressive performance in stereo matching, yet challenges remain due to sensor constraints, such as low dynamic range and motion blur. In contrast, event cameras capture per-pixel intensity changes asynchronously with high temporal resolution, making them less prone to motion blur and offering a high dynamic range. However, the event stream provides less spatial information compared to intensity images. Although existing state-of-the-art event-based stereo methods fuse features from both modalities, they still struggle to effectively capture and represent edge details in the scene. In this paper, we propose a novel edge-guided event-image stereo network, which utilizes extra edge cues to supplement edge information during disparity estimation. Firstly, we introduce an edge-guided event-image feature fusion approach to effectively supplement edge information in the fused features. Secondly, we incorporate edge cues into the disparity update process by introducing an edge-guided motion augmentation module, further augmenting the edge information during disparity estimation. Finally, we demonstrate the superiority of our method in stereo matching by conducting experiments on the real-world dataset using joint image and event data.

Existing event camera simulators primarily focus on the process of generating video events and often overlook the entire optical path in real-world camera systems. To address this limitation, we propose a novel Physical-based Event Camera Simulator (PECS), which is able to generate a high-fidelity realistic event stream by directly interfacing with the 3D scene. Our PECS features a lens simulation block for accurate light-to-sensor chip replication and a multispectral rendering module for precise photocurrent generation. We present two spatiotemporal event metrics to assess the similarity between simulated and actual camera events. Experimental results demonstrate that our PECS outperforms four state-of-the-art simulators by a large margin in terms of event-based signal fidelity. We integrate our PECS into the UE platform to generate extensive multi-task synthetic datasets and evaluate its effectiveness in downstream vision tasks (e.g., video reconstruction). Our open-source code can be available in the supplementary material.

The employment of the event-based synthetic aperture imaging (E-SAI) technique, which has the capability to capture high-frequency light intensity variations, has facilitated its extensive application on scene de-occlusion reconstruction tasks. However, existing methods usually require prior information and have strict restriction of camera motion on SAI acquisition methods. This paper proposes a novel end-to-end refocus-free variable E-SAI de-occlusion image reconstruction approach REDIR, which can align the global and local features of the variable event data and effectively achieve high-resolution imaging of pure event streams. To further improve the reconstruction of the occluded target, we propose a perceptual mask-gated connection module to interlink information between modules, and incorporate a spatial-temporal attention mechanism into the SNN block to enhance target extraction ability of the model. Through extensive experiments, our model achieves state-of-the-art reconstruction quality on the traditional E-SAI dataset without prior information, while verifying the effectiveness of the variable event data feature registration method on our newly introduced V-ESAI dataset, which obviates the reliance on prior knowledge and extends the applicability of SAI acquisition methods by incorporating focus changes, lens rotations, and non-uniform motion.

Recent advancements have achieved impressive results in removing Multi-Path Interference (MPI) and shot noise. However, these methods only utilize a single frame of ToF data, neglecting the correlation between frames. The multi-frame ToF denoising is still underexplored. In this paper, we propose the first learning-based framework for multi-frame ToF denoising. Different from previous frameworks, ours leverages the correlation between inter frames to guide the ToF noise removal with a confidence map. Specifically, we introduce a Dual-Correlation Estimation Module, which exploits both intra- and inter-correlation. The intra-correlation explicitly establishes the relevance between the spatial positions of geometric objects within the scene, aiding in depth residual initialization. The inter-correlation discerns variations in ToF noise distribution across different frames, thereby locating the areas with strong noise. To further leverage dual-correlation, we introduce a Confidence-guided Residual Regression Module to predict a confidence map, which guides the residual regression to prioritize the regions with strong ToF noise. The experimental evaluations have consistently shown that our approach outperforms other ToF denoising methods, highlighting its superior performance in effectively reducing strong ToF noise.

We seek to learn a generalizable goal-conditioned policy that enables zero-shot robot manipulation — interacting with unseen objects in novel scenes without test-time adaptation. While typical approaches rely on a large amount of demonstration data for such generalization, we propose an approach that leverages web videos to predict plausible interaction plans and learns a task-agnostic transformation to obtain robot actions in the real world. Our framework predicts tracks of how points in an image should move in future time-steps based on a goal, and can be trained with diverse videos on the web including those of humans and robots manipulating everyday objects. We use these 2D track predictions to infer a sequence of rigid transforms of the object to be manipulated, and obtain robot end-effector poses that can be executed in an open-loop manner. We then refine this open-loop plan by predicting residual actions through a closed loop policy trained with a few embodiment-specific demonstrations. We show that this approach of combining scalably learned track prediction with a residual policy requiring minimal in-domain robot-specific data enables zero-shot robot manipulation, and present a wide array of real-world robot manipulation results across unseen tasks, objects, and scenes.

We introduce DragAPart, a method that, given an image and a set of drags as input, generates a new image of the same object that responds to the action of the drags. Differently from prior works that focused on repositioning objects, DragAPart predicts part-level interactions, such as opening and closing a drawer. We study this problem as a proxy for learning a generalist motion model, not restricted to a specific kinematic structure or object category. We start from a pre-trained image generator and fine-tune it on a new synthetic dataset, Drag-a-Move, which we introduce. Combined with a new encoding for the drags and dataset randomization, the model generalizes well to real images and different categories. Compared to prior motion-controlled generators, we demonstrate much better part-level motion understanding.

In the realm of stochastic human motion prediction (SHMP), researchers have often turned to generative models like GANS, VAEs and diffusion models. However, most previous approaches have struggled to accurately predict motions that are both realistic and coherent with past motion due to a lack of guidance on the latent distribution. In this paper, we introduce Semantic Latent Directions (SLD) as a solution to this challenge, aiming to constrain the latent space to learn meaningful motion semantics and enhance the accuracy of SHMP. SLD defines a series of orthogonal latent directions and represents the hypothesis of future motion as a linear combination of these directions. By creating such an information bottleneck, SLD excels in capturing meaningful motion semantics, thereby improving the precision of motion predictions. Moreover, SLD offers controllable prediction capabilities by adjusting the coefficients of the latent directions during the inference phase. Expanding on SLD, we introduce a set of motion queries to enhance the diversity of predictions. By aligning these motion queries with the SLD space, SLD is further promoted to more accurate and coherent motion predictions. Through extensive experiments conducted on widely used benchmarks, we showcase the superiority of our method in accurately predicting motions while maintaining …

Generating human-object interactions (HOIs) is critical with the tremendous advances of digital avatars. Existing datasets are typically limited to humans interacting with a single object while neglecting the ubiquitous manipulation of multiple objects. Thus, we propose HIMO, a large-scale MoCap dataset of full-body human interacting with multiple objects, containing 3.3K 4D HOI sequences and 4.08M 3D HOI frames. We also annotate HIMO with detailed textual descriptions and temporal segments, benchmarking two novel tasks of HOI synthesis conditioned on either the whole text prompt or the segmented text prompts as fine-grained timeline control. To address these novel tasks, we propose a dual-branch conditional diffusion model with a mutual interaction module for HOI synthesis. Besides, an auto-regressive generation pipeline is also designed to obtain smooth transitions between HOI segments. Experimental results demonstrate the generalization ability to unseen object geometries and temporal compositions. Our data, codes, and models will be publicly available for research purposes.

Current approaches for 3D human motion synthesis generate high-quality animations of digital humans performing a wide variety of actions and gestures. However, a notable technological gap exists in addressing the complex dynamics of multi-human interactions within this paradigm. In this work, we present ReMoS, a denoising diffusion-based model that synthesizes full-body reactive motion of a person in a two-person interaction scenario. Assuming the motion of one person is given, we employ a combined spatio-temporal cross-attention mechanism to synthesize the reactive body and hand motion of the second person, thereby completing the interactions between the two. We demonstrate ReMoS across challenging two-person scenarios such as pair-dancing, Ninjutsu, kickboxing, and acrobatics, where one person’s movements have complex and diverse influences on the other. We also contribute the ReMoCap dataset for two-person interactions, containing full-body and finger motions. We evaluate ReMoS through multiple quantitative metrics, qualitative visualizations, and a user study, and also indicate usability in interactive motion editing applications. We will publicly release the code and the dataset with this paper to enable future research.

With the release of large-scale motion datasets with textual annotations, the task of establishing a robust latent space for language and 3D human motion has recently witnessed a surge of interest. Methods have been proposed to convert human motion and texts into features to achieve accurate correspondence between them. However, despite these efforts to align language and motion representations, we claim that the temporal element is often overlooked, especially for compound actions, resulting in chronological inaccuracies. To shed light on the temporal alignment in motion-language latent spaces, we propose Chronologically Accurate Retrieval (CAR) to evaluate the temporal understanding of the models. We decompose textual descriptions into events, and prepare negative text samples by shuffling the order of events in compound action descriptions. We then design a simple task for motion-language models to retrieve the more likely text between the ground truth and its chronologically shuffled version. CAR reveals many cases where current motion-language models fail to distinguish the event chronology of human motion, despite their impressive performance under conventional evaluation metrics. To achieve better temporal alignment between text and motion, we further propose to use these texts with shuffled sequences of events as negative samples to reinforce the motion-language models. …

This work introduces MotionLCM, extending controllable motion generation to a real-time level. Existing methods for spatial-temporal control in text-conditioned motion generation suffer from significant runtime inefficiencies. To address this issue, we first propose the motion latent consistency model (MotionLCM) for motion generation, building upon the latent diffusion model (MLD). By employing one-step (or few-step) inference, we further improve the runtime efficiency of the motion latent diffusion model for motion generation. To ensure effective controllability, we incorporate a motion ControlNet within the latent space of MotionLCM. This design enables explicit control signals to directly influence the generation process, similar to controlling other latent-free diffusion models for motion generation. By employing these techniques, our approach achieves real-time generation of human motion with text conditions and control signals. Experimental results demonstrate the remarkable generation and control capabilities of MotionLCM while maintaining real-time runtime efficiency.

We investigate exocentric-to-egocentric cross-view translation, which aims to generate a first-person (egocentric) view of an actor based on a video recording that captures the actor from a third-person (exocentric) perspective. To this end, we propose a generative framework called Exo2Ego that decouples the translation process into two stages: high-level structure transformation, which explicitly encourages cross-view correspondence between exocentric and egocentric views, and a diffusion-based pixel-level hallucination, which incorporates a hand layout prior to enhance the fidelity of the generated egocentric view. To pave the way for future advancements in this field, we curate a comprehensive exo-to-ego cross-view translation benchmark. It consists of a diverse collection of synchronized ego-exo tabletop activity video pairs from three public datasets: H2O, Aria Pilot, and Assembly101. The experimental results validate that Exo2Ego delivers photorealistic video results with clear hand manipulation details and outperforms several baselines in terms of both synthesis quality and generalization ability to new actions.

Speech sounds convey a great deal of information about the scenes, resulting in a variety of effects ranging from reverberation to additional ambient sounds. In this paper, we manipulate input speech to sound as though it was recorded within a different scene, given an audio-visual conditional example recorded from that scene. Our model learns through self-supervision, taking advantage of the fact that natural video contains recurring sound events and textures. We extract an audio clip from a video and apply speech enhancement. We then train a latent diffusion model to recover the original speech, using another audio-visual clip taken from elsewhere in the video as a conditional hint. Through this process, the model learns to transfer the conditional example's sound properties to the input speech. We show that our model can be successfully trained using unlabeled, in-the-wild videos, and that an additional visual signal can improve its sound prediction abilities.

Recent techniques for text-to-4D generation synthesize dynamic 3D scenes using supervision from pre-trained text-to-video models. However, existing representations for motion, such as deformation models or time-dependent neural representations, are limited in the amount of motion they can generate—they cannot synthesize motion extending far beyond the bounding box used for volume rendering. The lack of a more flexible motion model contributes to the gap in realism between 4D generation methods and recent, near-photorealistic video generation models. Here, we propose TC4D: trajectory-conditioned text-to-4D generation, which factors motion into global and local components. We represent the global motion of a scene bounding box using rigid transformation along a spline, and we learn local deformations that conform to the global trajectory using supervision from a text-to-video model. Our approach enables the synthesis of scenes animated along arbitrary trajectories, compositional scene generation, and significant improvements to the realism and amount of generated motion, which we evaluate qualitatively and through a user study.

Despite the recent progress in text-to-video generation, existing studies usually overlook the issue that only spatial contents but not temporal motions in synthesized videos are under the control of text. Towards such a challenge, this work presents a practical system, named LivePhoto, which allows users to animate an image of their interest with text descriptions. We first establish a strong baseline that helps a well-learned text-to-image generator (i.e., Stable Diffusion) take an image as a further input. We then equip the improved generator with a motion module for temporal modeling and propose a carefully designed training pipeline to better link texts and motions. In particular, considering the facts that (1) text can only describe motions roughly (e.g., regardless of the moving speed) and (2) text may include both content and motion descriptions, we introduce a motion intensity estimation module as well as a text re-weighting module to reduce the ambiguity of text-to-motion mapping. Empirical evidence suggests that our approach is capable of well decoding motion-related textual instructions into videos, such as actions, camera movements, or even conjuring new contents from thin air (e.g., pouring water into an empty glass). Interestingly, thanks to the proposed intensity learning mechanism, our system offers …

Image customization has been extensively studied in text-to-image (T2I) diffusion models, leading to impressive outcomes and applications. With the emergence of text-to-video (T2V) diffusion models, its temporal counterpart, motion customization, has not yet been well investigated. To address the challenge of one-shot motion customization, we propose Customize-A-Video that models the motion from a single reference video and adapting it to new subjects and scenes with both spatial and temporal varieties. It leverages low-rank adaptation (LoRA) on temporal attention layers to tailor the pre-trained T2V diffusion model for specific motion modeling from the reference videos. To disentangle the spatial and temporal information during the training pipeline, we introduce a novel concept of appearance absorbers that detach the original appearance from the single reference video prior to motion learning. The proposed modules are trained in a staged pipeline and inferred in a plug-and-play fashion, enabling easy extensions of our method to various downstream tasks such as custom video generation and editing, video appearance customization and multiple motion combination.

We present W.A.L.T, a diffusion transformer for photorealistic video generation from text prompts. Our approach has two key design decisions. First, we use a causal encoder to jointly compress images and videos within a unified latent space, enabling training and generation across modalities. Second, for memory and training efficiency, we use a window attention architecture tailored for joint spatial and spatiotemporal generative modeling. Taken together these design decisions enable us to achieve state-of-the-art performance on established video (UCF-101 and Kinetics-600) and image (ImageNet) generation benchmarks without using classifier free guidance. Finally, we also train a cascade of three models for the task of text-to-video generation consisting of a base latent video diffusion model, and two video super-resolution diffusion models to generate videos of 512 x 896 resolution at 8 frames per second.

This paper enables high-fidelity, transferable NeRF editing by frequency decomposition. Recent NeRF editing pipelines lift 2D stylization results to 3D scenes while suffering from blurry results, and fail to capture detailed structures caused by the inconsistency between 2D editings. Our critical insight is that low-frequency components of images are more multiview-consistent after editing compared with their high-frequency parts. Moreover, the appearance style is mainly exhibited on the low-frequency components, and the content details especially reside in high-frequency parts. This motivates us to perform editing on low-frequency components, which results in high-fidelity edited scenes. In addition, the editing is performed in the low-frequency feature space, enabling stable intensity control and novel scene transfer. Comprehensive experiments conducted on photorealistic datasets demonstrate the superior performance of high-fidelity and transferable NeRF editing. The code of our method will be made public.

We propose a simple but effective training-free method tailored to diffusion-based image-to-image translation. Our approach revises the original noise prediction network of a diffusion model by incorporating a noise correction term, which is based on the progressive interpolation of the textual prompts corresponding to a given source image and a desired target image. We formulate the noise correction term as the difference between two noise predictions; one is computed from the denoising network given a target latent and an adaptive interpolation between the source and target prompt embeddings, while the other is the noise prediction given the target latent and the source prompt embedding. The final noise prediction network is given by a combination of the standard denoising term and the noise correction term, where the firmer is designed to reconstruct must-be-preserved regions while the latter aims to effectively edit regions of interest relevant to the target prompt. Extensive experiments verify that the proposed method achieves outstanding performance with fast inference time and consistently improves existing frameworks when combined with them.

Diffusion models have made significant advances in text-guided synthesis tasks. However, editing user-provided images remains challenging, as the high dimensional noise input space of diffusion models is not naturally suited for image inversion or spatial editing. In this work, we propose an image representation that promotes spatial editing of input images using a diffusion model. Concretely, we learn to encode an input into image elements that can faithfully reconstruct an input image. These elements can be intuitively edited by a user, and are decoded by a diffusion model into realistic images. We show the effectiveness of our representation on various image editing tasks, such as object resizing, rearrangement, dragging, de-occlusion, removal, variation, and image composition.

Image stylization involves manipulating the visual appearance and texture (style) of an image while preserving its underlying objects, structures, and concepts (content). The separation of style and content is essential for manipulating the image's style independently from its content, ensuring a harmonious and visually pleasing result. Achieving this separation requires a deep understanding of both the visual and semantic characteristics of images, often necessitating the training of specialized models or employing heavy optimization. In this paper, we introduce B-LoRA, a method that leverages LoRA (Low-Rank Adaptation) to implicitly separate the style and content components of a single image, facilitating various image stylization tasks. By analyzing the architecture of SDXL combined with LoRA, we find that jointly learning the LoRA weights of two specific blocks (referred to as B-LoRAs) achieves style-content separation that cannot be achieved by training each B-LoRA independently. Consolidating the training into only two blocks and separating style and content allows for significantly improving style manipulation and overcoming overfitting issues often associated with model fine-tuning. Once trained, the two B-LoRAs can be used as independent components to allow various image stylization tasks, including image style transfer, text-based image stylization, consistent style generation, and style-content mixing.

We introduce Style Tailoring, a recipe to finetune Latent Diffusion Models (LDMs) in a distinct domain with high visual quality, prompt alignment and scene diversity. We choose sticker image generation as the target domain, as the images significantly differ from photorealistic samples typically generated by large-scale LDMs. We start with a competent text-to-image model, like Emu, and show that relying on prompt engineering with a photorealistic model to generate stickers leads to poor prompt alignment and scene diversity. To overcome these drawbacks, we first finetune Emu on millions of sticker-like images collected using weak supervision to elicit diversity. Next, we curate human-in-the-loop (HITL) Alignment and Style datasets from model generations, and finetune to improve prompt alignment and style alignment respectively. Sequential finetuning on these datasets poses a tradeoff between better style alignment and prompt alignment gains. To address this tradeoff, we propose a novel fine-tuning method called Style Tailoring, which jointly fits the content and style distribution and achieves best tradeoff. Evaluation results show our method improves visual quality by 14%, prompt alignment by 16.2% and scene diversity by 15.3%, compared to prompt engineering the base Emu model for stickers generation.

Creative processes such as painting often involve creating different components of an image one by one. Can we build a computational model to perform this task? Prior works often fail by making global changes to the image, inserting objects in unrealistic spatial locations, and generating inaccurate lighting details. We observe that while state-of-the-art models perform poorly on object insertion, they can remove objects and erase the background in natural images very well. Inverting the direction of object removal, we obtain high-quality data for learning to insert objects that are spatially, physically, and optically consistent with the surroundings. With this scalable automatic data generation pipeline, we can create a dataset for learning object insertion, which is used to train our proposed text-conditioned diffusion model. Qualitative and quantitative experiments have shown that our model achieves state-of-the-art results in object insertion, particularly for in-the-wild images. We show compelling results on diverse insertion prompts and images across various domains. In addition, we automate iterative insertion by combining our insertion model with beam search guided by CLIP.

For a given scene, humans can easily reason for the locations and pose to place objects. Designing a computational model to reason about these affordances poses a significant challenge, mirroring the intuitive reasoning abilities of humans. This work tackles the problem of realistic human insertion in a given background scene termed as Semantic Human Placement. This task is extremely challenging given the diverse backgrounds, scale, and pose of the generated person and, finally, the identity preservation of the person. We divide the problem into the following two stages i) learning semantic masks using text guidance for localizing regions in the image to place humans and ii) subject-conditioned inpainting to place a given subject adhering the scene affordance within the semantic masks. For learning semantic mask, we leverage rich object-scene priors learned from the text-to-image generative models and optimize a novel parameterization of the semantic mask, eliminating the need of large scale training. To the best of our knowledge, we are the first one to provide an effective solution for realistic human placements in diverse real world scenes. The proposed method can generate highly realistic scene compositions while perserving the background and subject identity. Further, we present results for several downstream …

In this paper we propose ProCreate, a simple and easy-to-implement method to improve sample diversity and creativity of diffusion-based image generative models and to prevent training data reproduction. ProCreate operates on a set of reference images and actively propels the generated image embedding away from the reference embeddings during the generation process. We collected a few-shot creative generation benchmark on eight different categories---encompassing different concepts, styles, and settings---in which ProCreate achieves the highest sample diversity and fidelity. Furthermore, we show that ProCreate is effective at preventing replicating training data in a large-scale evaluation using training text prompts.

Recent work has shown great progress in integrating spatial conditioning to control large, pre-trained text-to-image diffusion models. Despite these advances, existing methods describe the spatial image content using hand-crafted conditioning inputs, which are either semantically ambiguous (e.g., edges) or require expensive manual annotations (e.g., semantic segmentation). To address these limitations, we propose a new label-free way of conditioning diffusion models to enable fine-grained spatial control. We introduce the concept of neural semantic image synthesis, which uses neural layouts extracted from pre-trained foundation models as conditioning. Neural layouts are advantageous as they provide rich descriptions of the desired image, containing both semantics and detailed geometry of the scene. We experimentally show that images synthesized via neural semantic image synthesis achieve similar or superior pixel-level alignment of semantic classes compared to those created using expensive semantic label maps. At the same time, they capture better semantics, instance separation, and object orientation than other label-free conditioning options, such as edges or depth. Moreover, we show that images generated by neural layout conditioning can effectively augment real data for training various perception tasks.

This paper identifies significant redundancy in the query-key interactions within self-attention mechanisms of diffusion transformer models, particularly during the early stages of denoising diffusion steps. In response to this observation, we present a novel diffusion transformer framework incorporating an additional set of mediator tokens to engage with queries and keys separately. By modulating the number of mediator tokens during the denoising generation phases, our model initiates the denoising process with a precise, non-ambiguous stage and gradually transitions to a phase enriched with detail. Concurrently, integrating mediator tokens simplifies the attention module's complexity to a linear scale, enhancing the efficiency of global attention processes. Additionally, we propose a time-step dynamic mediator token adjustment mechanism that further decreases the required computational FLOPs for generation, simultaneously facilitating the generation of high-quality images within the constraints of varied inference budgets. Extensive experiments demonstrate that the proposed method can improve the generated image quality while also reducing the inference cost of diffusion transformers. When integrated with the recent work SiT, our method achieves a state-of-the-art FID score of 2.01. Source code will be released.

Recent advancements in text-to-image generative systems have been largely driven by diffusion models. However, single-stage text-to-image diffusion models still face challenges, in terms of computational efficiency and the refinement of image details. To tackle the issue, we propose CogView3, an innovative cascaded framework that enhances the performance of text-to-image diffusion. CogView3 is the first model implementing relay diffusion in the realm of text-to-image generation, executing the task by first creating low-resolution images and subsequently applying relay-based super-resolution. This methodology not only results in competitive text-to-image outputs but also greatly reduces both training and inference costs. Our experimental results demonstrate that CogView3 outperforms SDXL, the current state-of-the-art open-source text-to-image diffusion model, by 77.0\% in human evaluations, all while requiring only about 1/2 of the inference time. The distilled variant of CogView3 achieves comparable performance while only utilizing 1/10 of the inference time by SDXL.

We propose Context Diffusion, a diffusion-based framework that enables image generation models to learn from visual examples presented in context. Recent work tackles such in-context learning for image generation, where a query image is provided alongside context examples and text prompts. However, the quality and context fidelity of the generated images deteriorate when the prompt is not present, demonstrating that these models are unable to truly learn from the visual context. To address this, we propose a novel framework that separates the encoding of the visual context and the preservation of the desired image layout. This results in the ability to learn from the visual context and prompts, but also from either one of them. Furthermore, we enable our model to handle few-shot settings, to effectively address diverse in-context learning scenarios. Our experiments and human evaluation demonstrate that Context Diffusion excels in both in-domain and out-of-domain tasks, resulting in an overall enhancement in image quality and context fidelity compared to counterpart models.

In recent years, deep generative models have developed rapidly and can generate high-quality images based on input texts. Assessing the quality of generated images in a way consistent with human preferences is critical for both generative model evaluation and preferred image selection. Previous works aligned models with human preferences by training scoring models on image pairs with preference annotations (e.g., ImageReward and HPD). These carefully annotated image pairs well describe human preferences for choosing images. However, current training paradigm of these preference models is to directly maximize the preferred image score while minimizing the non-preferred image score in each image pair through cross-entropy loss. This simple and naive training paradigm mainly has two problems: 1) For image pairs of similar quality, it is unreasonable to blindly minimize the score of non-preferred images and can easily lead to overfitting. 2) The human robustness to small visual perturbations is not taken into account, resulting in the final model being unable to make stable choices. Therefore, we propose Stable Preference to redefine the training paradigm of human preference model and a anti-interference loss to improve robustness to visual disturbances. Our method achieves state-of-the-art performance on two popular text-to-image human preference datasets. Extensive ablation …

Text-to-image diffusion models (SD) exhibit significant advancements while requiring extensive computational resources. Existing acceleration methods usually require extensive training and are not universally applicable. LCM-LoRA, trainable once for diverse models, offers universality but rarely considers ensuring the consistency of generated content before and after acceleration. This paper proposes SpeedUpNet (SUN), an innovative acceleration module, to address the challenges of universality and consistency. Exploiting the role of cross-attention layers in U-Net for SD models, we introduce an adapter specifically designed for these layers, quantifying the offset in image generation caused by negative prompts relative to positive prompts. This learned offset demonstrates stability across a range of models, enhancing SUN's universality. To improve output consistency, we propose a Multi-Step Consistent (MSC) loss, which stabilizes the offset and ensures fidelity in accelerated content. Experiments on SD v1.5 show that SUN leads to an overall speedup of more than 10 times compared to the baseline 25-step DPM-solver++, and offers two extra advantages: (1) training-free integration into various fine-tuned Stable-Diffusion models and (2) state-of-the-art FIDs of the generated data set before and after acceleration guided by random combinations of positive and negative prompts.

Text-to-image diffusion models are a class of deep generative models that have demonstrated an impressive capacity for high-quality image generation. However, these models are susceptible to implicit biases that arise from web-scale text-image training pairs and may inaccurately model aspects of images we care about. This can result in suboptimal samples, model bias, and images that do not align with human ethics and preferences. In this paper, we present an effective scalable algorithm to improve diffusion models using Reinforcement Learning (RL) across a diverse set of reward functions, such as human preference, compositionality, and fairness over millions of images. We illustrate how our approach substantially outperforms existing methods for aligning diffusion models with human preferences. We further illustrate how this substantially improves pretrained Stable Diffusion (SD) models, generating samples that are preferred by humans 80.3% of the time over those from the base SD model while simultaneously improving both the composition and diversity of generated samples.

With the ability to generate high-quality images, text-to-image (T2I) models can be exploited for creating inappropriate content. To prevent misuse, existing safety measures are either based on text blacklists, easily circumvented, or harmful content classification, using large datasets for training and offering low flexibility. Here, we propose Latent Guard, a framework designed to improve safety measures in text-to-image generation. Inspired by blacklist-based approaches, Latent Guard learns a latent space on top of the T2I model's text encoder, where we check the presence of harmful concepts in the input text embeddings. Our framework is composed of a data generation pipeline specific to the task using large language models, ad-hoc architectural components, and a contrastive learning strategy to benefit from the generated data. Our method is evaluated on three datasets and against four baselines.

This paper presents Arc2Face, an identity-conditioned face foundation model, which, given the ArcFace embedding of a person, can generate diverse photo-realistic images with an unparalleled degree of face similarity than existing models. Despite previous attempts to decode face recognition features into detailed images, we find that common high-resolution datasets (e.g. FFHQ) lack sufficient identities to reconstruct any subject. To that end, we meticulously upsample a significant portion of the WebFace42M database, the largest public dataset for face recognition (FR). Arc2Face builds upon a pretrained Stable Diffusion model, yet adapts it to the task of ID-to-face generation, conditioned solely on ID vectors. Deviating from recent works that combine ID with text embeddings for zero-shot personalization of text-to-image models, we emphasize on the compactness of FR features, which can fully capture the essence of the human face, as opposed to hand-crafted prompts. Crucially, text-augmented models struggle to decouple identity and text, usually necessitating some description of the given face to achieve satisfactory similarity. Arc2Face, however, only needs the discriminative features of ArcFace to guide the generation, offering a robust prior for a plethora of tasks where ID consistency is of paramount importance. As an example, we train a FR model on synthetic …

Significant advancements have been achieved in the domain of face generation with the adoption of diffusion models. However, diffusion models tend to amplify biases during the generative process, resulting in an uneven distribution of sensitive facial attributes such as age, gender, and race. In this paper, we introduce a novel approach to address this issue by debiasing the attributes in generated images. Our approach involves disentangling facial attributes by localizing the means within the latent space of the diffusion model using Gaussian mixture models (GMM). This method, leveraging the adaptable latent structure of diffusion models, allows us to localize the subspace responsible for generating specific attributes on-the-fly without the need for retraining. We demonstrate the effectiveness of our technique across various face datasets, resulting in fairer data generation while preserving sample quality. Furthermore, we empirically illustrate its effectiveness in reducing bias in downstream classification tasks without compromising performance by augmenting the original dataset with fairly generated data.

As Diffusion Models have shown promising performance, a lot of efforts have been made to improve the controllability of Diffusion Models. However, how to train Diffusion Models to have the disentangled latent spaces and how to naturally incorporate the disentangled conditions during the sampling process have been underexplored. In this paper, we present a training framework for disentangling the latent spaces of Diffusion Models. We further propose two sampling methods that can boost the realism of our Diffusion Models and also enhance the controllability. Concisely, we train Diffusion Models conditioned on two latent features, a spatial content mask, and a flattened style embedding. We rely on the inductive bias of the denoising process of Diffusion Models to encode pose/layout information in the content feature and semantic/style information in the style feature. Regarding the sampling methods, we first extend Composable Diffusion Models by breaking the conditional independence assumption to allow for some dependence between conditional inputs, which is shown to be effective in realistic generation in our experiments. Second, we propose timestep-dependent weight scheduling for content and style features to further improve the performance. We also observe better controllability of our proposed methods compared to existing methods in image manipulation and …

Recent advancements in diffusion-based generative image editing have sparked a profound revolution, reshaping the landscape of image outpainting and inpainting tasks. Despite these strides, the field grapples with inherent challenges, including: i) inferior quality; ii) poor consistency; iii) insufficient instrcution adherence; iv) suboptimal generation efficiency. To address these obstacles, we present ByteEdit, an innovative feedback learning framework meticulously designed to Boost, Comply, and Accelerate Generative Image Editing tasks. ByteEdit seamlessly integrates image reward models dedicated to enhancing aesthetics and image-text alignment, while also introducing a dense, pixel-level reward model tailored to foster coherence in the output. Furthermore, we propose a pioneering adversarial and progressive feedback learning strategy to expedite the model's inference speed. Through extensive large-scale user evaluations, we demonstrate that ByteEdit surpasses leading generative image editing products, including Adobe, Canva, and MeiTu, in both generation quality and consistency. ByteEdit-Outpainting exhibits a remarkable enhancement of 388% and 135% in quality and consistency, respectively, when compared to the baseline model. Experiments also verfied that our acceleration models maintains excellent performance results in terms of quality and consistency.