04Section
Solutions
The landscape of models, methods and tools for football intelligence, mapped and judged. Seven families, one question put to each: what does it tell us about who has the ball?
- 7families
- 47solutions judged
- 74models, methods and tools reviewed
- 3ruled out by a licence or a rule
Where our own system sits
Everything on this map is someone else’s method, read for what it tells us about who has the ball. Our own system runs one of them: an exact Viterbi decode per half, over football evidence scored on every frame. The design is drawn in The Shrewdies decoder, and its result against the organisers’ nearest-player baseline is on the Initial Challenge page.
How to read this map
Each family owns one link of the chain above, and each family page explains how its methods work, how to choose between them, what they cannot do and where our decoder stands. Every entry answers the same questions: how it works, where it wins, what it costs, how it plugs into possession. Where a licence or a competition rule put a tool out of reach, the entry says so under “Our call”. Three do. The calls that shaped our own design are in Architecture.
The families
-
Diagram: Per-frame answers, Possession, Home, Away, Dead ball. 01 of 07
Direct possession
Methods that answer the question itself: which team has the ball on this frame. From the organisers' nearest-player rule to a graph model that needs no ball at all.
- Best task fit on the map: the output is the label.
- The rules go silent when the ball does.
- The learned models need labels and identities most feeds lack.
6 entries
-
Diagram: Home arrives first, Away arrives first. 02 of 07
Physics and features
Distances, arrival times, pitch control, filtered trajectories. Evidence computed from physics, with nothing trained and every number traceable.
- Label-free and deterministic.
- Turns who is closest into who gets there first.
- Evidence, never the answer: controlling space is not having the ball.
6 entries
-
Diagram: Home, Away, Dead ball, Best path. 03 of 07
Classifiers and decoders
Models that turn evidence into probabilities, and decoders that turn probabilities into possessions. Home of Viterbi, the algorithm our decoder is built on.
- Classifiers need labels. Decoders need only evidence.
- Frame accuracy does not buy coherent possessions.
- The algorithm our decoder runs.
11 entries
-
Diagram: Detect, Track, Calibrate, Missed ball, Identity switch, Metres off. 04 of 07
Vision and calibration
Detectors, trackers and calibrators that rebuild players, ball and camera from video. The route to positions when a tracking feed is missing or has holes.
- Detect, track, calibrate: three stages, one error budget.
- The ball is the hardest object on the pitch to see.
- The most varied licences on the map.
10 entries
-
Diagram: ±1 s, 25 frames, Spotted events, Possession between. 05 of 07
Events and replay
Spotters that timestamp passes, drives and tackles, and the tools that untangle a broadcast edit. Every possession change is an event; not every event is a change.
- Events anchor possession changes.
- Scored at one second: 25 frames of doubt.
- Silent between touches, which is most of the match.
7 entries
-
Diagram: Video clock, Tracking clock, Aligned. 06 of 07
Synchronisation
Putting video, audio and tracking on one clock. Invisible when right. When wrong, every boundary moves and every comparison is between two different moments.
- Ten frames off moves every boundary by 0.4 s.
- Check the file's own clock first.
- On a pitch, projected player positions are the best clock.
5 entries
-
Diagram: Provider file, Parser, One model, Coordinates, Dataframe. 07 of 07
Readers and formats
Libraries that load provider files into one model and pin events to frames. The tooling under every cross-provider test. The datasets themselves live in the Datasets section.
- One loader across providers.
- Events pinned to tracking frames.
- Formats are standardised. Meanings are not.
2 entries
Start here
-
The exact best path through per-frame evidence. Three states, a cost for every switch, and a guarantee that no better sequence exists. Our evidence decoder is built on it.
- Sequence
- Exact
- No training
- Fielded
-
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
-
Who would reach each point of the pitch first. The physical model behind most spatial football analytics, in its reference tutorial code and a fast implementation.
- Tracking
- Physics
- No training
- MIT
-
The strongest general-purpose learners for tabular data. Give each frame a row of distances, speeds and visibility flags and they return H/A/D probabilities.
- Feature table
- Learned
- Needs labels
- MIT / Apache-2.0
-
The full chain from broadcast video to a minimap: detect, track, identify team, role and shirt number, and place every athlete on the pitch.
- Video
- Full pipeline
- Open weights
- GPL-3.0
Find a page
All 47 pages
- Nearest-player baseline
- PathCRF
- Possession TCN (Borghesi et al.)
- Rule-based event detection (Vidal-Codina et al.)
- Match-phase segmentation
- DEFCON
- Pitch control
- Accessible space (DAS)
- Kalman filter and RTS smoother
- ruptures
- floodlight
- unravelsports
- Viterbi decoding
- Hidden Markov models (hmmlearn)
- Hidden semi-Markov models (hsmmlearn)
- Linear-chain CRF (pytorch-crf)
- Gradient-boosted trees (LightGBM, XGBoost)
- Set encoders (Deep Sets, Set Transformer)
- Temporal convolutional networks
- Temporal action segmentation (MS-TCN, ASFormer)
- Graph network toolkits (PyG, PyG Temporal, DGL)
- Bayesian online change-point detection
- Learned imputation (BRITS, PyPOTS)
- Roboflow Sports: soccer detectors
- Game State Reconstruction baseline
- Tracking by detection (ByteTrack, BoT-SORT, StrongSORT)
- SportsLabKit
- SoccerNet re-identification
- SAM 2
- CoTracker
- PnLCalib (and No Bells, Just Whistles)
- TVCalib
- KpSFR
- T-DEED and its team-aware variant
- Ball-action spotting, 2023 winning solution
- E2E-Spot
- FOOTPASS baselines
- SoccerNet spotting development kits
- PySceneDetect
- SoccerNet replay grounding
- FFprobe and PyAV
- Audio offset estimation (GCC-PHAT)
- Dynamic time warping (dtw-python)
- Geometric synchronisation
- Learned video alignment (TCC, TPL)
- kloppy
- DataBallPy