Research questionHow can we estimate reliable correspondences between partially observed 3D shapes under strong non-isometric deformation?Missing regions and strong non-isometric deformation make it difficult to identify which points on two 3D shapes correspond. The challenge is to infer accurate matches without relying on representations that are costly or generalize poorly. Latest papersRecent research connected to this question, newest first.TokenMatch: 3D Mesh Correspondence Transformer with Curvature-Guided TokenisationThe source presents TokenMatch, a feed-forward transformer trained only on BeCoS for partial-to-partial matching and evaluated on CP2P, PSMAL, BeCoS, FAUST, SCAPE, and SHREC'19. It reports mean geodesic error, intersection-over-union, sub-second inference, and generalization to full shapes without retraining or fine-tuning.research paper · Sep 3, 2026