Self-Geometry method boosts vision model accuracy using epipolar constraints at test time

Researchers have introduced Self-Geometry, a test-time adaptation pipeline that enforces multi-view geometric consistency in vision foundation models without requiring full retraining. Unlike prior methods that rely on a model's own outputs for self-consistency, Self-Geometry uses pseudo ground-truth 2D correspondences optimized against epipolar and multi-view losses. The approach was evaluated across six vision foundation models and four benchmark datasets, consistently improving both pose accuracy and depth scores. On the challenging ETH3D benchmark, the method improved pose accuracy of the VGGT model by up to 37.3% on a key metric. Despite running in under two minutes per scene, the pipeline is not suited for real-time use and may struggle in texture-poor or dynamic environments where reliable correspondences are hard to obtain.