Category-Level 3D Correspondence in Camera Space via Morphable Object Priors

Researchers introduced HouseCorr3D, a new benchmark with 178,000 images across 50 household object categories and 3D keypoint annotations on CAD models.

The dataset includes occlusion-aware and symmetry labels, making it a richer testbed for category-level 3D correspondence learning.

They also proposed Morpheus, which learns a shared morphable shape prior by separating canonical shape, deformation, and pose.

This lets semantically consistent 3D correspondences emerge in camera space from single images, with state-of-the-art results on HouseCorr3D.

The work could improve robotics and AR/VR by enabling better part matching across object instances.