SeeSE3: Emergence of 3D Space in Vision Features

arXiv cs.AI · 6 hours ago

Researchers investigated whether vision foundation models develop internal representations that correspond to 3D Euclidean space structure by proposing topological and geometric probes to test the alignment between feature spaces and SE(3) transformations. Self-supervised vision models showed strong correlation with 3D space when probed with a mutual neighborhood metric and Poincaré Adapter, despite having no explicit 3D supervision during training. This finding enabled development of latent-space navigation techniques that perform visual odometry and localization directly in feature space without reconstructing explicit 3D geometry.