02Reference

Glossary and failure map

The vocabulary of possession inference in one place, and the seven false equivalences that sink most first attempts.

  • Data
  • Geometry
  • Decoding
  • Scoring
Five families of terms, one problem. Most mistakes come from treating two neighbouring cells as the same thing.

Seven false equivalences

It is tempting to assume The truth
Closest player = possessor Proximity flips in every pass and duel; control does not
Touch = turnover Possession moves when the other side establishes control
Ball untracked = dead ball The tracker loses the ball in live play, at its most contested
Smoother labels = better spells Smoothing deletes real short possessions and merges their neighbours
More models = more evidence Models fed by the same tracking share the same errors
Valid homography = synchronised frames Where on the pitch and when in the match are separate questions
Published code = usable model A repository without weights, licence and data path is a paper

Possession and scoring

Term Meaning
Controlled possession A team has the ball until the other side establishes control or play stops
H, A, D Home in possession, away in possession, dead ball
X Abstention in a submission; never a true state
Possession spell An unbroken run of H or A; a stoppage ends it
spellF1 Ranking metric: one-to-one matching of spells by temporal IoU, averaged over five thresholds
macroF1 Tie-break: mean frame-level F1 over H, A and D
Temporal IoU Overlap of two intervals divided by their union
False split, false merge One true spell predicted as two; two predicted as one
Pareto frontier The candidates no other candidate beats on both metrics
Leakage Using the answer, or anything derived from it, as an input

Data and geometry

Term Meaning
ATD Automatic Tracking Data: positions of players and ball for every video frame
Track slot A column of tracking output; a team, not a player identity
Tracklet A short reliable run of one track before the object is lost
Tactical camera The wide view of the whole pitch; the competition’s reference clock
Broadcast camera The edited feed: live shots, cuts, close-ups and replays
Smart Tagging Automatic tactical intervals and restarts; context, not an event log
Homography Plane-to-image projective transform; no height, no timing
Provenance Where an observation came from, and whether it was measured or inferred

Time and decoding

Term Meaning
Timebase The clock a source is indexed on
Dense temporal classification A label for every frame, as opposed to sparse event timestamps
Observation model Turns tracking, tags or pixels into per-frame evidence for each state
Decoder Turns a sequence of noisy evidence into one coherent state sequence
Viterbi Exact dynamic-programming search for the best path under switch costs
HMM, HSMM Hidden Markov model; its semi-Markov form also models how long states last
CRF A discriminative model of whole label sequences
TCN Temporal convolutional network: learned features over a window of time
DTW Dynamic time warping: aligns two sequences that run at different rates
Filtering, smoothing Estimating from the past only; estimating with the future as well

Learning on players

Term Meaning
Permutation invariance The output does not change when the players are listed in another order
Deep Sets An order-invariant network over a group of players
Set Transformer Attention over unordered objects
GNN Graph neural network: players and ball as nodes, interactions as edges
Ball-free inference Reading possession from player movement alone

The methods behind these terms are mapped and judged in Solutions; the data they learn from is in Datasets.

Sources

  1. CHALLENGE (organisers' repository)github.com
  2. DATA (organisers' repository)github.com
  3. EVALUATION (organisers' repository)github.com