Estimate Optical Flow (Triton)

How it fits

Typical backends

Caveats

• State lifetimeA frame deque is held per component instance across input messages and is lost only on container restart; the first frame returns a zero-filled flow tensor because there is no neighbour yet, which downstream must distinguish from a real stationary frame.
• I/O contractAn empty input frame is skipped (not buffered) but still emits a zero flow tensor sized to that frame; the model output is bilinearly resized from the model's internal resolution to the source frame, so detail beyond the model resolution is interpolated, not measured.
• Parameter interactionwith_respect_to_second_frame flips the sign convention of the emitted vectors, so consumers that assume one sign must pin it at backend authoring time.
• Parameter interactionframe_diff is rejected at zero; larger values widen the temporal baseline (retaining more buffered frames) and amplify slow motion at the cost of noise.
• Parameter interactionoutput_layout must match how the served network emits its flow map: 'channels_first' for two contiguous [H, W] planes (the common [2, H, W] export) and 'channels_last' for an interleaved [H, W, 2] map. A mismatch silently scrambles the two flow components.
• CompatibilityInference runs against a model-serving service that must be reachable; cold start includes the model load before the first flow tensor is produced.
• LatencyEach frame is a full neural forward pass on the serving backend, so the achievable frame rate is bounded by that service's throughput.