Skip to content
arXiv cs.AI · Papers

Triangular Consistency as a Universal Constraint for Learning Optical Flow

arXiv:2606.19938v3 Announce Type: replace-cross Abstract: We propose triangular consistency as a first-principled constraint for optical flow, which is agnostic to network architecture, supervision type, and dataset, and applies to both image-pair and multi-frame settings. This simple but powerful constraint is to comp