03Datasets family
Frame-level possession labels
The scarce resource. Three releases put a possession field on the tracking itself, frame for frame. Only one labels home, away and dead ball under an open licence.
- 3datasets with possession on the tracking
- 81matches between them
- 7open matches with all three states
What this family is
The only data where the answer is in the file. A possession field sits on the tracking, frame for frame, so a model’s output can be laid against it and scored. Every other family on this site gets near possession. These three get to it.
That is why the family is small. Labelling possession on every frame of a match is expensive, the definition is a provider’s own, and the footage behind it is rarely free. Three public releases exist. They do not agree with each other, and that disagreement is the first thing to understand.
Three providers, three grammars
| Dataset | What the label says | Dead ball | Tracking behind it |
|---|---|---|---|
| DFL IDSSE | Team and ball status | Explicit | Official league tracking, 25 Hz |
| Gradient Sports | Owning team and ball state | Explicit | Broadcast-derived |
| SkillCorner | A possession field | Not labelled | Broadcast-derived, about 10 Hz |
How to choose
| You need | Use | Because |
|---|---|---|
| A verdict on a method | DFL IDSSE | Three states, official data, CC BY 4.0. Anyone can rescore it |
| Breadth for a sequence model | Gradient Sports | 64 matches of tournament football in one schema |
| To study missing evidence | SkillCorner | Every position says whether it was seen or extrapolated |
Judge on one provider, learn on another. A model that only agrees with the labels it grew up on has learned a vendor, not the game.
The traps
- Semantics. Possession is each provider’s definition. Mapping it to home, away and dead ball is a modelling decision. Make it per provider and keep it in the open.
- Dead ball is where they differ most. SkillCorner labels who, never whether the ball is alive. A gap between phases is not a stoppage.
- Frame rate. 10 Hz hides short touches that 25 Hz shows. Upsampling invents precision.
- Camera type. Broadcast-derived tracks include players the camera never saw. Do not pool them with league tracking without a provider flag.
- Licences. Only DFL is CC BY 4.0. Gradient comes by request under publisher terms. SkillCorner’s MIT licence is on the repository.
- Leakage. Gradient and SoccerNet-GAR are cut from the same 64 matches. Split by match across both.
What it cannot do
It cannot make providers stop mattering. Eighty-one matches across three definitions is not a corpus you can pour into one loss function. And none of the three ships video or tactical tags, so the family tests the tracking half of a pipeline only.
Where it meets our decoder
This family chose our submission. The competition rules forbid tuning on the match’s own answers, so the decision had to be made elsewhere, on labels we did not make, and DFL was the only open data with all three states.
We damaged a clean DFL match until it looked like the competition feed, with the ball hidden in bursts on 41% of frames and 0.15 m of position noise, and let it choose between the decoder’s parameter profiles. Nothing was trained on it. The decoder has no trained weights.
The scores are on the DFL IDSSE page. How the search ran is on the Pareto lens.
In this family3
-
Seven Bundesliga and 2. Bundesliga matches with official 25 Hz tracking and a possession and ball-status flag on every frame. The open judge for any possession model.
- Tracking
- Possession labels
- CC BY 4.0
-
Tracking, events and rosters for all 64 matches of the 2022 World Cup, with an owning team and a ball state in the tracking. The largest possession-labelled release.
- Tracking
- Possession labels
- 64 matches
- By request
-
Ten A-League matches of broadcast-derived tracking with a possession field, phases of play and, crucially, a flag that says which positions were seen and which were extrapolated.
- Tracking
- Possession field
- Detected or extrapolated