03Events
SoccerNet Action Spotting
More than 500 complete broadcast matches with 17 classes of timestamped events: fouls, throw-ins, corners, goals. The largest open view of how football stops and restarts.
- Broadcast video
- 17 event classes
- 500+ matches
- Modality
- Match-length broadcast video with timestamped events
- Scale
- 500+ complete matches
- Classes
- 17, including foul, throw-in, corner, offside and goal
- Companion labels
- Camera shots and replays
- Possession labels
- None
- Calibration
- None per frame
- Licence
- SoccerNet research terms, non-commercial. Video under NDA
- Access
- Labels public (SN-Labels). Video by NDA
What it is
The task SoccerNet is best known for, in its 17-class SoccerNet-v2 form. Whole broadcast matches, each annotated with the moments that structure a game: the ball going out, throw-ins, fouls and the free kicks that follow, corners, offsides, substitutions, cards, goals. Companion annotations mark camera shots and replays.
What it teaches
| Signal | What you get |
|---|---|
| Possession | The dead-ball state, by its edges: what stops play and what restarts it |
| Tracking | Nothing |
| Events | Seventeen classes over full matches |
| Calibration | Nothing |
Where it wins
- Full matches at scale. Hundreds of complete games, so the rhythm of stoppages is observed, not sampled.
- Restart vocabulary. Throw-in, corner, free kick, kick-off: the events that end a dead ball.
- Broadcast grammar. Replay and camera labels teach a model when the picture is not showing live play at all.
What it costs
- Instants, not durations. A foul has a timestamp. When the ball actually went dead, and when it came alive again, is not annotated.
- Sparse. Open play is almost unmarked.
- Rights. The footage is NDA-gated and non-commercial.
How it plugs into possession
Dead ball is the state with the least evidence. The ball is often off-screen or untracked, players walk, the camera cuts to a replay. Positions alone say little. Restart events say a lot: a throw-in fixes the moment the ball came back alive and usually which team has it.
That is the role tactical tags play in our decoder. Restart windows are evidence for the dead-ball state before them and for possession after them. This dataset is where a video model can learn to produce the same kind of cue.
Our read
A dead-ball dataset in disguise. Most of its classes are a stoppage or a restart, which is the third possession state and the one most models get wrong.