Announcement Diysc

New pre-print: “Do It Yourself: Learning Semantic Correspondence from Pseudo-Labels” is now available on arXiv! We show that foundational features can be refined with an adapter that is trained with pseudo-labels, which are themselves zero-shot predictions using the same foundational features. We improve the quality of pseudo-labels through 3D-aware chaining with cycle-consistency and reject wrong pairs using a spherical prototype. New state-of-the-art results on SPair71k and scalable to larger datasets. Accepted to ICCV 2025!