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

Name: Teaching DINOv3 About Partial 3D Geometry: A Self-Supervised Geometry-Aware Approach
Uploaded: 2026-10-20T13:15:33.700312+00:00
Duration: 214 s
Description: GeoLoRA is a self-supervised method that refines DINOv3 features by injecting 3D awareness using LoRA adaptation. It generates robust feature descriptors for partial 3D shapes by optimizing alignment between full and partial views using a geodesic-based contrastive loss, leading to state-of-the-art performance in partial shape matching.

GeoLoRA is a self-supervised method that refines DINOv3 features by injecting 3D awareness using LoRA adaptation. It generates robust feature descriptors for partial 3D shapes by optimizing alignment between full and partial views using a geodesic-based contrastive loss, leading to state-of-the-art performance in partial shape matching.

Open the full episode