Stereo matching and surface normal estimation provide complementary geometric cues, yet existing methods often model them independently and remain unreliable in low-light, reflective, and transparent regions. We present GeoStereo, a unified stereo geometry estimation framework that jointly predicts disparity and surface normals. GeoStereo couples a feed-forward stereo branch with a diffusion-based normal branch through a shared feature extractor, disparity-to-normal initialization, and aligned cross-view guidance. This design allows diffusion priors to improve disparity estimation in ill-posed regions, while stereo geometry provides structured guidance for normal prediction. Extensive zero-shot experiments demonstrate strong generalization across challenging scenes. GeoStereo achieves Rank-1 disparity performance on multiple benchmarks, including KITTI and NYUv2, and obtains the best normal estimation error on several real indoor datasets.
Overview of our proposed framework. A stereo pair is first processed by a shared feature extractor to obtain shared features, which are aggregated to predict disparity and further refined by a ConvGRU. The initial disparity is converted into a disparity-to-normal initialization and, together with the shared left-view features and the warp-to-left-view condition, is fed into a diffusion-based U-Net. The latent is iteratively refined and decoded to produce the final normal map.
GeoStereo couples feed-forward stereo matching with diffusion-based surface normal estimation under a shared feature extractor, allowing both tasks to exchange complementary geometric priors.
The initial disparity is converted into a coarse normal initialization, providing a structured starting point for diffusion denoising while preserving consistency with stereo geometry.
The right image is differentiably warped into the left-view coordinate system using predicted disparity, supplying aligned binocular cues for normal refinement in occluded and ambiguous regions.
The diffusion branch contributes strong structural priors in low-light, reflective, transparent, and weak-texture regions, improving both disparity robustness and fine-grained surface geometry.
Qualitative comparison of zero-shot inference on diverse scenes. All methods are trained on a mixture of public datasets. The examples include challenging cases such as low-light conditions, reflections, and transparent objects. The top three rows show disparity results, while the bottom row shows surface normal predictions.
