03Datasets family
Broadcast video and game state
Television pictures turned into players, teams and touches. This is the upstream problem: what a possession model consumes is what these benchmarks ask vision to produce.
- 4datasets
- 102,992ball actions tied to a player
- 87,939moments in both video and tracking
What this family is
Broadcast footage, annotated to the level a possession model could feed on: boxes, track IDs, positions on the pitch, team and jersey, and the touches themselves. If there is no tracking feed, this is how one gets made.
It is also the family with the hardest camera. A broadcast director follows the ball, cuts to close-ups and runs replays. A tactical camera does none of that.
What each one annotates
| Dataset | Unit | Annotates | Access |
|---|---|---|---|
| Game State Reconstruction | Clips | Pitch position, role, team, jersey number | Labels open, video NDA |
| SoccerNet Tracking | Clips | Boxes and track IDs, ball included | Labels open, video NDA |
| SoccerNet-GAR | 87,939 event windows | Ten activity classes, video and tracking in sync | NDA |
| FOOTPASS | 54 full matches | 102,992 ball actions, each with its player | CC BY-NC 4.0, video NDA |
How to choose
| You need | Use |
|---|---|
| The full chain from pixels to pitch positions | Game State Reconstruction |
| An honest grade for a ball detector | SoccerNet Tracking |
| To ask whether video adds anything to tracking | SoccerNet-GAR |
| Who touched the ball, and when | FOOTPASS |
From touches to possession
Two of the four carry real possession signal. SoccerNet Tracking is where you measure how the ball is lost: how often, for how long, and in which scenes. FOOTPASS shows who played it. A run of actions by one team is a spell. A change of acting team is a turnover somewhere between two timestamps.
The traps
- Part of the pitch. Off-screen players do not exist in the labels. A back line can be missing for a whole attack.
- Clips break identity. Track IDs hold inside a clip and mean nothing across clips. Only a jersey number survives a cut.
- Rights. SoccerNet video sits behind an NDA. FOOTPASS annotations are non-commercial. Nothing trained here ships freely.
- Leakage. SoccerNet-GAR shares its 64 matches with the Gradient Sports release.
What it cannot do
Give a spell. Touches are instants and boxes are positions. Nothing in the family marks how long a team held the ball, or when it went dead and came back.
Where it meets our decoder
Our Initial Challenge decoder begins where this family ends: it decodes the automatic tracking the organisers supplied.
Two lessons from this family are in the design all the same. Ball dropouts come in bursts, so our validation hid the ball in bursts. And a touch is stronger evidence than proximity: in the decoder a reception can change possession, and one frame of a nearer opponent cannot. The upstream chain itself is mapped in Video-only inference.
In this family4
-
Broadcast clips where every player and referee is placed on the pitch with a role, a team and a jersey number. The benchmark for turning television pictures into tracking data.
- Broadcast video
- Pitch localisation
- Team and jersey
-
The standard multi-object tracking benchmark for broadcast football: boxes and track IDs for players, referees and the ball in crowded 1080p clips.
- Broadcast video
- Boxes
- Track IDs
- Ball
-
87,939 group activities from all 64 World Cup 2022 matches, each one available twice: as broadcast video and as player and ball tracking for the same moment.
- Video and tracking
- 64 matches
- NDA
-
54 full professional matches with 102,992 ball actions, each one anchored to the player who made it. The densest open record of who touched the ball, and when.
- Broadcast video
- Actor-linked actions
- CC BY-NC 4.0