Teaching DINOv3 About Partial 3D Geometry: A Self-Supervised Geometry-Aware Approach

Figure 1. GeoLoRA (Geometry-Aware Low-Rank Adaptation) injects geometry awareness into 2D foundation features for partial shape matching. Our method is self-supervised and produces robust features for different kinds and amounts of partiality, even for out-ofdistribution real-world shapes (in the image, from left: two shapes cut by a plane; two partial views; a full template for reference; three shapes with holes; a statue from [1]).