3 ms·

WAFT: Warping-Alone Field Transforms for Optical Flow (2025)

• peter_d_sherman 2mo ago
  >"Optical flow is a fundamental low-level vision task that estimates per-pixel 2D motion between video frames. It has many downstream applications, including 3D reconstruction and synthesis [...] action recognition [...] frame interpolation [...] and autonomous driving..." [...] >"WAFT is similar to RAFT but replaces cost volume with high-resolution warping, achieving better accuracy with lower memory cost." [...] >"Compared with existing methods, WAFT ranks 1st on Spring, Sintel, and KITTI benchmarks, achieves the best zero-shot generalization on KITTI, while being 1.3 − 4.1× faster than existing methods that have competitive accuracy (e.g., 1.3× than Flowformer++, 4.1× than CCMR+)."