Estimate Optical Flow (Classical)

How it fits

Typical backends

Caveats

• State lifetimeA deque of recent grayscale frames is held in the running process; until frame_diff + 1 frames have been seen (and on any frame whose size differs from the buffered ones) an all-zero (H, W, 2) tensor is emitted and the buffer reseeds. A restart clears the buffer.
• I/O contractInput is converted to grayscale before flow computation, so chroma-only motion is ignored; an empty Image emits an empty (0, 0, 2) tensor without buffering the frame. Output is always at the source frame resolution.
• Parameter interactionmethod selects the algorithm and which knobs apply: the dis_* params only affect DIS, and pyr_scale, levels, winsize, iterations, poly_n, poly_sigma only affect Farnebäck. TV-L1 takes no extra knobs. DIS is the default because it is both faster and more accurate than Farnebäck.
• Parameter interactionSetting both model_width and model_height downscales both frames before flow, then resizes the field back to source and rescales the x and y components by source_width/model_width and source_height/model_height. Leaving them at 0 runs at native resolution.
• Parameter interactionwith_respect_to_second_frame (default true) gives forward flow; false swaps the frame pair and negates the result for backward flow. frame_diff widens the temporal gap between the two compared frames.
• AccuracyAll three classical methods are weaker than neural flow on large displacements; prefer a learned estimator when fast-motion accuracy matters.
• State lifetimeAll parameters are captured once at startup; runtime changes have no effect and require a redeploy.