Geometry Grounding: Elevating Blind Distortion Correction with 3D Structural Priors

Geometric distortion breaks the projective mapping and ge-ometric relations in images, which degrades the performance and relia-bility of computer vision tasks. Although deep network-based methodsfor single-image distortion correction have achieved remarkable progress,most existing approaches optimize only for 2D pixel-level alignment, suchas optical flow losses or image reconstruction errors. Without explicit 3Dgeometric constraints, such models may produce visually plausible recti-fications yet fail to enforce 3D-consistent structure. To address this lim-itation, this paper proposes a 3D geometry guided framework for singleimage distortion correction, named Geometry-Grounded Distortion Cor-rection (G2DC). G2DC introduces a pretrained 3D foundation model toextract multi-dimensional geometric representations as supervision, andincorporates projective geometry priors into feature learning. Specifically,G2DC adds a geometric constraint scheme and jointly optimizes a struc-tural alignment loss and a physical alignment loss. The former minimizesfeature mismatch to ensure that the corrected output is structurally closeto the undistorted scene. The latter enforces strict 3D constraints usingthe predicted point map, depth topology, and camera parameter regular-ization, so the model preserves 3D structure consistency and plausibilitywhile restoring 2D appearance. Experiments on several widely used dis-tortion benchmark datasets show that G2DC outperforms state-of-the-art methods, especially on metrics that are sensitive to 3D geometricconsistency. Project webpage: https://gitee.com/VR_NAVE/g2dc.git.