Matching image regions to 3D shapes without texture

This work addresses the challenge of establishing correspondence between 2D image segments and 3D shape parts—a problem complicated by differences in appearance, geometry, and viewpoint. The approach distills features from a 2D vision model onto a 3D shape surface, then identifies "Best Segmentation Buddies": vertices whose most similar image pixels lie within the corresponding image segmentation region. This enables reliable cross-modal matching. The method works directly with untextured 3D shapes and requires only image segmentation labels, demonstrating robustness across diverse real-world image-shape pairs. Code and interactive results are available on the project page.