FreeFlow: A Bias-free Hierarchical Transformer for Optical Flow Estimation
Conventional optical flow estimation has long relied on task-specific heuristics, such as explicit correlation volumes, feature warping, and iterative refinement recurrent loops. FreeFlow departs from this established pattern by discarding specialized inductive biases entirely in favor of a general hierarchical transformer architecture. Built as a standard feed-forward encoder-decoder, it balances local and long-range context through windowed, shifted-window, and downscaled global attention. The design delivers state-of-the-art accuracy across standard benchmarks—including Sintel and KITTI-2015—while scaling predictably with capacity and maintaining memory efficiency at 1080p resolution.
Why it's worth reading
It demonstrates that dense optical flow estimation can achieve state-of-the-art accuracy with a clean, general-purpose transformer, bypassing decades of hand-crafted architectural heuristics.