04Solutions family
Direct possession
Methods that answer the question itself: which team has the ball on this frame, or is play dead. Six entries, two routes, one shared blind spot.
- 6entries
- 2need no labels
- 1works without the ball
One picture: a frame in, an answer out
Every method in this family takes tracking and says home, away or dead. The difficulty is in the picture above: answers given frame by frame flicker, and possession does not. What separates the entries is where their answer comes from.
There are two routes. Rules read possession off geometry: the nearest player, a touch, a ball crossing the line. Learned models read it off examples. A network sees labelled frames by the thousand and learns what possession looks like, with the ball or without it.
The other families each own one link of the possession chain: reading, seeing, syncing, measuring, spotting, decoding. A direct method replaces the measuring, spotting and decoding links with one step. That is its strength, since there is nothing to assemble. It is also its risk: when it fails, there is no intermediate evidence to inspect.
How to choose inside it
| Entry | Labels | Ball and identities | Compute |
|---|---|---|---|
| Nearest-player baseline | None | Ball required. No identities | Trivial |
| Rule-based event detection | None | Ball required. No identities | Trivial |
| PathCRF | To train or retrain | No ball. Persistent player tracks | GPU to train; saved models exist |
| Possession TCN | Frame labels, many matches | As trained. No identities | GPU to train |
| Match-phase segmentation | Phase labels | Possession given as input | GPU to train |
| DEFCON | Its own training data | Possession and players known | GPU to train |
Read the table as a ladder. With no labels, start with the rules. With labels and clean identities, PathCRF is the one entry that survives a missing ball. The last two rows are not ways to find possession. They are what gets built once it is found.
What this family cannot do
- Rules cannot see past the ball. Both rule-based entries are keyed on ball position. No ball, no answer.
- Learned models cannot outrun their labels. They learn one provider’s definition of possession on one provider’s tracking. A new feed means new labels.
- Most cannot promise a coherent sequence. PathCRF can, because it carries a decoder inside. The others need one behind them.
Where our decoder stands
It is a direct possession method by output, and a member of the classifiers and decoders family by construction. It keeps the baseline’s cue, a player within 1.5 m of the ball and closer than any opponent. It treats an untracked ball as no evidence. An exact Viterbi decode chooses the sequence, and no labels are fitted. The design is in The Shrewdies decoder.
In this family6
-
The organisers' reference method: give the ball to whoever stands closest, then vote over a window. It sets the bar and shows exactly where proximity stops being possession.
- Tracking
- Ball-dependent
- Rule-based
- No training
-
Possession inferred from the players alone. A dynamic player graph, set attention and a CRF recover who holds the ball and where it travels, with no ball coordinates in the input.
- Tracking
- Ball-free
- Learned
- MPL-2.0
-
The closest published relative of the H/A/D task: a permutation-invariant player encoder feeding a temporal convolutional network, trained to say who has the ball from tracking.
- Tracking
- Learned
- Needs labels
- Paper only
-
Passes, shots and receptions recovered from tracking by physical rules: who is close enough to the ball, when its motion changes, when it leaves play.
- Tracking
- Ball-dependent
- Rule-based
- Paper only
-
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
-
Graph neural networks that credit defenders for what they prevent. Possession is its input, not its output, which is exactly why it belongs on the map.
- Tracking
- Graph
- Learned
- Needs possession