04Vision and calibration

SoccerNet re-identification

The benchmark and baseline for recognising the same player across camera views, in a sport where every team-mate wears the same shirt.

  • Video
  • Identity
  • Benchmark
  • MIT
Candidates are ranked cue by cue. In football the kit column never separates team-mates.
Modality
Player crops from several broadcast cameras
Output
Identity embeddings and match rankings
Evidence
Official SoccerNet task with challenge baselines
Code and weights
Development kit and baseline training code
Licence
MIT code; data under SoccerNet's terms

Re-identification asks whether two image crops show the same person. In surveillance the cue is clothing. In football clothing is the one thing that cannot help: ten outfield players share a kit, and the task becomes telling them apart by build, hair, boots, posture and a shirt number when it happens to be visible.

SoccerNet’s task frames this across cameras. A player seen in one view of an action must be matched to the same player seen from another angle. The development kit provides the data loaders, a baseline built on standard re-identification training, and the ranking metrics.

Where it wins

  • Football-specific. Models trained here learn the cues that survive identical kits.
  • Cross-view matching is what multi-camera and replay analysis need.
  • Its embeddings can replace pedestrian models in appearance-based trackers.

What it costs

  • Low-resolution crops limit what any model can see. At wide-shot scale, team-mates are near-identical.
  • Matching within one action is the easy case. Identity across a whole match is harder.
  • Embeddings say “same” or “different”, not who. Naming a player takes shirt-number reading or a roster.

How it plugs into possession

Team possession does not need player identity. Player possession does, and so does everything built on it: who lost the ball, who won it, pass networks. Re-identification is how anonymous tracks become players.

It also guards the team label. The strongest appearance cue is kit colour, and an appearance embedding separates the two teams far more reliably than it separates team-mates. Used as a team classifier, it is working on the easy version of its own problem.

Our read

Identity is the hard residue of football vision. Not needed for H/A/D, indispensable the moment the question becomes who.

Sources

  1. sn-reid: development kit for the SoccerNet re-identification challengegithub.com