Production Status — May 2026

Computer vision for Program A football

Every label PFF gives you — formation, coverage, personnel, tendencies — derived automatically from raw broadcast video. 3,833 plays trained, 63 games processed, real model predictions on real broadcast frames.

Headline Accuracy

Coverage on labeled NFL tracking

88.3%

Macro accuracy across all 5 coverage shells

C0: 94% · C1: 83% · C2: 90% · C3: 89% · C4: 85%

2,293 held-out NFL plays (different games from training, from BDB 2021+2022 data)

Coverage on Program A/Program B broadcast

94.5%

Macro within trained-on games

C0: 97% · C1: 91% · C2: 92% · C3: 94% · C4: 98%

1,942 Program A/Program B broadcast plays processed through full pipeline (yolov8l → ByteTrack → homography → coord extraction)

Formation on labeled NFL tracking

97.3%

6 formations, all ≥90%

SHOTGUN: 98% · EMPTY: 98% · SINGLEBACK: 96% · I_FORM: 90% · PISTOL: 98% · JUMBO: 100%

3,500+ held-out NFL plays

Pipeline Health

Player detection (broadcast)

95%

yolov8l detects 21/22 players at conf≥0.25 on Program A broadcast (yolov8n was 13/22 = 59%)

Homography fire rate

100%

up from 50%

Broadcast plays extracted

1,942

end-to-end pipeline

NFL tracking training set

15,794

BDB plays w/ PFF labels

Demo Surfaces

LIVE VISION

Watch the model run

Pick a real Program A/Program B play. Snap frame + 4-second clip with field-coord overlay. Shows the model identifying formation + coverage shell with real confidence scores.

Open →

SCOUT REPORTS

Tendency + matchup analytics

Live tendency views derived from 244 PFF columns × 3,833 plays. Personnel, formation, coverage, situational breakdowns for Program A and Program B.

Open →

NATURAL LANGUAGE

Ask the data anything

Chat over the labeled play data. 'What did Program B run on third-and-long in 2024?' Returns analytics + the source plays. Same query against vision-derived labels works identically.

Open →

Honest Framing

94.5% on broadcast is the operational deployment number — coaches label some plays in a game, model predicts the rest with that accuracy. For brand-new opponents the model has never seen labels from, accuracy drops to ~35%. Closing that gap requires more game diversity in training (conference film exchange, additional PFF charting, or our weak-supervision pipeline over public NFL film).