04Events and replay
FOOTPASS baselines
Player-centric ball-action spotting: not only what happened and when, but which player did it, using tracks, a graph over players and a denoising sequence model.
- Video + tracks
- Player-level events
- Identity-dependent
- Apache-2.0 code
- Modality
- Broadcast video with player tracks
- Output
- Actions attributed to individual players
- Evidence
- Baselines for SoccerNet's 2026 player-centric challenge
- Code and weights
- Training and inference code
- Licence
- Apache-2.0 code; data and models may be non-commercial
Event spotting says a pass happened at 12:31. These models say who played it. They are the reference baselines of SoccerNet’s 2026 player-centric ball-action challenge, built on the FOOTPASS dataset, where each action has to be assigned to a specific player. That is the form event data takes when providers sell it.
The baselines combine three views of the game. Track-aware action detection looks at each tracked player and scores what he is doing. A graph network relates players to one another, since a pass has a sender and a receiver. A sequence model then reads the resulting stream of candidate actions and denoises it using tactical regularity. Play-by-play football has a grammar: a reception follows a pass.
Where it wins
- Player-level attribution, the granularity of commercial event data.
- It uses the structure of play itself as a prior, not only pixels.
- It combines video and tracking-style inputs in one model.
- The code is Apache-2.0.
What it costs
- It needs to know who is who. Attribution to players presumes reliable identities, and wrong identities corrupt the output directly.
- A deep stack of trained components, with the data and compute that implies.
- Data and model resources may be non-commercial even though the code is not.
- Player actions are not team possession labels. A mapping step remains.
How it plugs into possession
A player-attributed action stream is nearly a possession sequence already. Each action names the player on the ball, and his team follows. Where identities exist, output of this kind is the richest event evidence a possession decoder can receive.
Its denoising stage is the same idea as sequence decoding, learned: use what football allows to correct what the detector saw. The difference is the unit. Here it is actions. In a possession decoder it is frames.
The Initial Challenge feed carries a team and no identity, so team-level evidence and a team-level decode fit it as it is: Players without names.
Our read
Where event spotting is going: actions with names attached. It needs identities, and it shows what they are worth.