Partial shape matching—aligning or matching 3D shapes that are only partially observed—is a key challenge in 3D vision, especially for real-world applications where full views are rare. Despite its importance, the problem has received less attention than full-shape matching.

Existing feature descriptors, which are commonly used for shape matching, struggle when parts of the geometry are missing. This limits their effectiveness in practical scenarios.

The paper presents a self-supervised, geometry-aware approach that teaches DINOv3, a vision foundation model, to handle partial 3D geometry. By incorporating geometric awareness into the training process, the method aims to improve feature robustness for partial shapes.

The work is described in an arXiv preprint; no experimental results are provided in the abstract, so the effectiveness of the approach remains to be seen.