The Play-Detection Cascade

The system does not rely on a single learned model. Instead it fuses cheap, interpretable signals with increasing precision:

1. Field Mask — computed once per source. All subsequent motion analysis ignores everything outside the field lines.
2. Camera Motion Compensation — modern global motion estimation subtracts operator movement so that player motion isn't masked by camera work (the primary failure mode in 2013 research).
3. Active-Cell Motion (VTI) — divides the masked frame into an N×N grid. Each cell is classified active/inactive across a sweep of thresholds. Real events (snap, tackle pile-up, whistle) produce consistent peaks across many thresholds. This is viewpoint- and zoom-invariant.
4. Audio Onset — detects whistle peaks and crowd reaction. Serves as confirmation rather than primary gate (broadcast mix varies).
5. Formation Check — low-fps player detection (permissive models only) verifies the pre-snap geometry that distinguishes "about to start a play" from random motion or huddles.

Candidate intervals are fused into start/end timestamps with an explicit confidence score (0–1). Clips are emitted with asymmetric padding: -2.5 s before snap, +2.0 s after end. This padding was measured to reduce bad cuts from 63% (snap-timed) to 8%.

Any clip whose confidence falls below the review threshold is placed in a coach-facing human review queue. The product philosophy is that reviewing 8 uncertain boundaries out of 110 plays is acceptable; a silently wrong cut is not.