FreeFlow: A Bias-free Hierarchical Transformer for Optical Flow Estimation
FreeFlow reports state-of-the-art optical flow results without using the usual flow-specific machinery.
The paper replaces correlation volumes, warping, and iterative refinement with a feed-forward hierarchical transformer. Its attention stack mixes local window attention, shifted windows, and reduced-resolution global attention. The authors report benchmark results of 0.68/1.48 EPE on Sintel Clean/Final, 3.23 Fl-all on KITTI-2015, and 3.192 1px on Spring. They also say the model scales cleanly from small to large variants and stays memory efficient at 1080p inference. HF Daily Papers' note
The paper replaces correlation volumes, warping, and iterative refinement with a feed-forward hierarchical transformer. Its attention stack mixes local window attention, shifted windows, and reduced-resolution global attention. The authors report benchmark results of 0.68/1.48 EPE on Sintel Clean/Final, 3.23 Fl-all on KITTI-2015, and 3.192 1px on Spring. They also say the model scales cleanly from small to large variants and stays memory efficient at 1080p inference. HF Daily Papers' note
score 5