Compute dense optical flow between frames

Use This When

• Tracking motion in scenes without distinct objects or with complex deformations
• Building motion-based segmentation or anomaly detection where pixel-level flow matters
• Analyzing scene dynamics, camera motion, or crowd flow patterns
• Feeding motion context to downstream trackers or stabilization logic

What It Does

• Buffers video frames and computes dense per-pixel motion vectors using RAFT-style deep network
• Outputs 2-channel flow field matching input frame dimensions with horizontal and vertical motion at each pixel
• Configurable frame spacing controls temporal range vs latency tradeoff
• Can anchor flow to fixed reference frame or compute frame-to-frame differences

Works Best With

• Any video source → this component → motion visualizers, object movement extractors, or trackers
• Pipelines needing full-field motion without relying on feature matching or detections
• Workflows combining flow with segmentation to get per-object motion summaries

Caveats

• Frame spacing creates latency; frame_diff=5 at 30fps means 5-frame buffer delay
• Flow accuracy degrades with fast motion, occlusions, or lighting changes between frames
• Requires Triton-served RAFT model; ensure model preprocessing params match training expectations