A Unified Stereo Geometry Estimation Framework for Disparity and Surface Normal

GeoStereo is a hybrid stereo vision framework that jointly estimates disparity and surface normals.

It combines feed-forward disparity prediction with a diffusion-based normal branch so the two tasks reinforce each other.

The model uses disparity-to-normal initialization and warped left-view conditioning to improve 3D geometry estimation in hard scenes.

It shows stronger performance across benchmarks such as KITTI, NYUv2, iBims-1, and ScanNet, especially in low-light, reflective, and transparent regions.