03Datasets family
Tracking and full-pitch video
Positions without possession. These four teach what a tactical camera sees and how far a position can be trusted, which is the noise floor under every possession feature.
- 4datasets
- 15.5 hfull-pitch video under CC BY
- 2matches with sensor-measured positions
What this family is
Players on a pitch, frame after frame, with no word on who has the ball. For possession the value is indirect and large. Every possession feature is a distance between positions: who is nearest the ball, who is within playing distance. The quality of the positions is the ceiling on the quality of the answer.
The family covers both ends of how tracking is made: the full-pitch video going in, and the pitch coordinates coming out.
What is actually measured
| Dataset | Measured | Ball | Licence |
|---|---|---|---|
| Metrica sample data | Both teams in pitch coordinates, with events | Yes | None stated |
| SoccerTrack v2 | Players on the pitch plane from panoramic video, with identity | Derived from ball actions | CC BY 4.0 |
| TeamTrack | Player boxes and track IDs in 4K to 8K video | No | MIT code, data terms apart |
| Alfheim | Home players by wearable sensor, plus panoramic video | No | Non-commercial research |
How to choose
| You need | Use |
|---|---|
| To learn Metrica’s file and coordinate conventions | Metrica sample data |
| To train or test a tracker in the tactical view | SoccerTrack v2 |
| To stress identity through crowded play | TeamTrack |
| To measure localisation error in metres | Alfheim |
The traps
- Camera type. This is full-pitch footage. Players are small in frame, and a detector trained on broadcast close-ups does not transfer.
- Where the ball comes from. In SoccerTrack v2 the ball may be interpolated between annotated actions. It is a label derived from events, not a measurement.
- Identity is not team. A track ID is not a player. When a tracker swaps two opponents, the team of the “nearest player” flips with it.
- Coordinates. Metrica’s positions are normalised to the unit square. No distance means anything until a pitch size is chosen.
- Licences. Alfheim forbids re-identification and performance profiling. TeamTrack’s code licence does not cover its data.
What it cannot do
Name the ball’s owner. None of the four carries a possession label, and two have no ball at all. Metrica’s events give a sparse skeleton of who played the ball, which is enough to sanity-check a decoder and too thin to grade one.
Where it meets our decoder
The competition tracking is this family’s hard case: automatic, team-only, with no player identities. Our decoder reads control as a player within 1.5 m of the ball and closer than any opponent, then makes every change of state pay a cost, so one noisy frame, or one swapped track, cannot flip possession.
Position noise was part of our validation for the same reason: the DFL match we judged on was degraded with 0.15 m of noise before scoring. See Players without names for the identity problem in full.
In this family4
-
Three anonymised matches of tracking with synchronised events, in the file conventions Metrica uses everywhere. The format bridge to any Metrica-shaped problem.
- Tracking
- Events
- Open repository
-
Ten full matches filmed as one panoramic full-pitch view, with pitch-space player tracks, identities and ball actions. The best openly licensed stand-in for a tactical camera.
- Full-pitch video
- Tracking
- CC BY 4.0
-
A full-pitch multi-object tracking benchmark in 4K to 8K, with millions of annotated player boxes from side and top cameras. Built to break trackers on identity.
- Full-pitch video
- Player boxes
- Track IDs
-
Two professional matches filmed by a multi-camera panoramic rig while the home players wore position sensors. Video and an independent measurement of where the players really were.
- Panoramic video
- Wearable sensors
- Research only