Dense Correspondence of Deformable Shapes with Different Identities

Establish dense correspondences between 3D deformable shapes that have different identities and undergo non-rigid deformations, including cases in which one or both shapes are only partially observed.

Background

The paper identifies dense correspondence estimation as a fundamental task in 3D vision because it supports information transfer and the learning of statistical models. The unresolved setting involves geometries with different identities, such as distinct humans, that also undergo non-rigid deformations, such as changes in pose.

The problem becomes more difficult when one or both shapes are partial observations caused by occlusions, acquisition noise, or local reconstruction errors. The proposed GeoLoRA method addresses partial shape matching experimentally, but the broader correspondence problem is presented as remaining unresolved.