Revisiting Cross-View Completion: Self-Supervised Pre-Training via Reconstruction Error Comparison

Gekko uses relative reconstruction error between cross-view and masked-autoencoder predictions as a self-supervised co-visibility proxy, adding binocular training signals that consistently outperform CroCo on 3D vision tasks while training directly from raw video.

Thibaut Loiseau, Guillaume Bourmaud, Vincent Lepetit

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026