Geometry Matters: 3D Foundation Priors for Learning Semantic Correspondence

Researchers introduced a post-training framework for semantic correspondence that injects reconstructed 3D geometry into 2D foundation features.

The method combines DINO and Stable Diffusion features with SAM3D geometry, renders PartField descriptors, and filters weak matches using geodesic distance.

A lightweight adapter is then trained on the cleaned correspondences, improving accuracy over prior post-training methods without pose annotations or spherical approximations.

It helps resolve hard cases like left-right ambiguity and repeated parts while reducing the need for manual supervision.