The number first.
You decide the rest.
Primary signals
- Person detection
- Jersey-number OCR
- Your roster
- Team and burst context
Reference safeguards
- Off until your organisation enables it
- Limited to your organisation and roster
- Single appearance signals never auto-publish
How a photo is processed
- 01
Detect people
Every frame is scanned for players. A crowd shot with nobody in focus resolves to nothing and costs you no review time.
- 02
Find the jersey region
The area of the shirt likely to carry a number is isolated — front, back or angled.
- 03
Read the number
OCR tuned for the shapes that actually go wrong: 1 and 7, 5 and 6, 8 and 0, and a leading digit hidden by an arm.
- 04
Match the roster
The number is searched only within the teams attached to the event. Seven is ambiguous across a league and near-unique inside one match.
- 05
Combine independent signals
When a number is hidden, clear facial identity, matching kit, several reference views, the event roster and person-level burst tracking can corroborate one another.
- 06
Assign or ask
Only an unambiguous multi-signal result is assigned automatically. A single face, similar kit, pose or loose resemblance stays in review.
Your corrections matter, but they do not retrain anything
Every correction is captured as structured feedback: what was predicted, what you chose, the model version, and the frame it came from. None of it retrains a model automatically. It remains inside the organisation as reviewable correction data. A player can be excluded from AI assignment, which also deletes uploaded references and derived appearance data for that player.
Start with the next shoot