The Pipeline — From Film to Football Intelligence

The master film file is treated as immutable. We ingest it once; all subsequent analysis, restoration, clipping, and annotation produce *derivatives* that always resolve back to the original frame numbers and pixels. This decision prevents silent drift in labels over time.

"Foggy" or low-quality film is rarely atmospheric fog. It is almost always bitrate starvation combined with crushed blacks and limited dynamic range.

We diagnose using metrics calibrated on real football film (200 endzone + 200 sideline frames). The primary signal is bits-per-pixel — it discriminates 17× between clean and heavily degraded sources. Blockdetect was thoroughly tested and discarded: it returned 2.826 on clean film versus 2.811 on badly degraded film, making it useless as a quality metric. Blurdetect showed similarly poor discrimination.

Restoration follows a classical, non-generative ladder keyed to these measurements:

1. Black-level correction
2. Contrast stretch (skipped when dynamic range span already exceeds 0.88 — real football film typically hits ~0.91)
3. Targeted denoising
4. Sharpening

The resulting enhanced video is a *viewing derivative only*. It is never fed into computer-vision training. All labels and training data always reference the original master pixels (Decision D2 from the design spec).

Detecting play boundaries in continuous broadcast or NFHS film is harder than academic literature assumes. We use a cascaded, multi-signal approach rather than a single model:

• Field mask — restricts all analysis to the playing surface, eliminating crowd and sideline noise.
• Global camera motion compensation — removes pan/tilt/zoom so player motion can be isolated.
• Active-cell Variable Threshold Image (VTI) — zoom-invariant motion measure adapted from IAAI-13 research. Counts "active" grid cells across multiple thresholds to find robust peaks for snap and play-end.
• Audio onset detection — whistle and crowd reaction as independent confirmation.
• Formation geometry check — RF-DETR player detection confirms the characteristic pre-snap "two opposing rows, low motion, ~22 bodies."

Measured performance: cutting exactly at the estimated snap produced 63% bad cuts. Adding 2 s of pre-roll reduced bad cuts to 8%. Every proposed boundary receives a confidence score. Boundaries below threshold route to a human review queue in the coaching app rather than risking silent errors.

Every label carries full provenance: source film, model/version, confidence score, and schema version. The football concept taxonomy provides consistent terminology for formations, responsibilities, and concepts.

The resulting corpus is strictly first-party or properly licensed film only. No scraped or terms-of-service-violating sources are ingested.