Triangular Consistency Constraint Improves Optical Flow Across Architectures

Composing two optical flows to enforce consistency with a third — across image pairs, video frames, or synthetic transforms — yields a plug-and-play training constraint with negligible compute overhead and no extra annotations. The method is architecture- and supervision-agnostic, showing consistent improvements across tested settings.