04Direct possession
Match-phase segmentation
Graph-temporal models that cut a match into tactical phases such as build-up, progression and finishing. One level above possession, and dependent on it.
- Tracking
- Graph + sequence
- Learned
- Preprints
- Modality
- Tracking, with possession as context
- Output
- Tactical phase per frame or window
- Evidence
- Two 2026 preprints: seven Bundesliga matches; 203 matches
- Code and weights
- Papers only
- Licence
- Not applicable: no released code
Two 2026 preprints attack the same problem from the same direction. Possession says who has the ball. Phase says what they are doing with it.
The first segments in-possession play by intention, into phases such as build-up, progression and finishing. A temporal graph attention network encodes player relations frame by frame and a Transformer models how they evolve. It is evaluated on seven Bundesliga matches.
The second pairs a graph network with an LSTM to classify seven match phases across 203 matches. Its most useful contribution is an ablation of the graph itself: Delaunay adjacency against other topologies, with spatial attributes on the edges. The authors report gains over feature-based baselines.
| Intention-driven phases | GNN-LSTM phases | |
|---|---|---|
| Spatial model | Temporal graph attention | Graph network, Delaunay edges |
| Temporal model | Transformer | LSTM |
| Target | In-possession phases | Seven match phases |
| Evaluation | Seven Bundesliga matches | 203 matches |
Where it wins
- Phases are the vocabulary coaches use. This is the layer where tracking becomes tactics.
- Both papers show that relations between players, not positions alone, carry the tactical signal.
- The Delaunay result is practical. Which graph you draw matters as much as which network you run on it.
What it costs
- Both take possession as given. The first segments only in-possession play. The second feeds possession and ball status to the model as input features.
- Phase labels are expensive expert annotation, and definitions differ between analysts and clubs.
- Neither paper ships code or weights.
How it plugs into possession
After possession, never before. A phase model trained with possession as an input cannot be turned round to infer possession: the answer is already in the question.
Used in order, the stack is clean. Possession inference labels each frame H, A or D. A phase model subdivides the H and A stretches into what the team was trying to do.
The architectural lesson does transfer. A graph encoder per frame with a sequence model across frames is the shape of PathCRF, and the shape any supervised possession model built on graph networks would take.
Our read
The layer above possession. Valuable once H/A/D is solved, circular if used to solve it.