Estimate Optical Flow (PTLFlow)

How it fits

Typical backends

Caveats

• State lifetimeRecent frames are held per component instance across messages and reset only on container restart or a mid-stream resolution change; the first frame (and any frame fewer than frame_diff steps after a reset) returns a zero-filled flow tensor because there is no comparable earlier frame yet, which downstream must distinguish from a real stationary frame.
• I/O contractInference runs at the model resolution set by model_height and model_width, but the flow is resized back to the source frame and the vectors are rescaled per axis, so the emitted OpticalFlow is always at the source frame resolution and aligns pixel-for-pixel with full-resolution masks; setting both to zero runs natively and skips the rescale.
• Hard constraintmodel_variant must name a model the installed ptlflow version provides, and ckpt_path must name a checkpoint alias that model defines; an unknown model name or alias aborts startup.
• Parameter interactionmodel_variant trades speed for accuracy across the zoo and iters adds refinement steps at a latency cost on RAFT-family models only (non-recurrent models ignore it); frame_diff widens the temporal baseline; with_respect_to_second_frame flips flow direction; device, model_height, model_width, and with_respect_to_second_frame are re-read each frame and can be retuned live.
• LatencyThe checkpoint downloads and loads on cold start, so first-frame latency includes the load time; RAFT-family models route correlation through a compiled CUDA kernel and the component is GPU-oriented.