Non-Identical Object Matching

Conventional feature matching looks for correspondences between images of the same object or scene captured under different viewpoints or conditions. Humans, however, can naturally connect objects that aren’t identical but are semantically related — a cheetah and a husky, a photograph and an illustration.

We develop models for this kind of semantic correspondence between non-identical objects, including matching across different individuals or species of wildlife — a theme with direct real-world applications in areas like wildlife monitoring.

Progress so far

We presented matching across different wildlife species at the CVPR 2024 workshop (CV4A), and are presenting a method for matching semantically similar non-identical objects at WACV 2026.