360° Image Perception with MLLMs: A Comprehensive Benchmark and a Training-Free Method

Multimodal Large Language Models (MLLMs) have shownimpressive abilities in understanding and reasoning over conventionalimages. However, their perception of 360° images remains largely un-derexplored. Unlike conventional images, 360° images capture the en-tire surrounding environment, enabling holistic spatial reasoning butintroducing challenges such as geometric distortion and complex spa-tial relations. To comprehensively assess MLLMs’ capabilities to per-ceive 360° images, we introduce 360Bench, a Visual Question Answering(VQA) benchmark featuring 7K-resolution 360° images, seven represen-tative (sub)tasks with annotations carefully curated by human annota-tors. Using 360Bench, we systematically evaluate seven MLLMs and sixenhancement methods, revealing their shortcomings in 360° image per-ception. To address these challenges, we propose Free360, a training-free scene-graph-based framework for high-resolution 360◦ VQA. Free360decomposes the reasoning process into modular steps, applies adaptivespherical image transformations to 360° images tailored to each step,and seamlessly integrates the resulting information into a unified graphrepresentation for answer generation. Experiments show that Free360consistently improves its base MLLM and provides a strong training-free solution for 360◦ VQA tasks. The source code and dataset will beavailable at https://tranhuyen1191.github.io/360Bench-Free360/.