03The data atlas

Datasets

The football-data landscape, read through one question: what can this dataset teach a possession model?

Four kinds of evidence. Only one layer says who has the ball.
  • 22datasets mapped
  • 6families
  • 3carry possession on the tracking
  • 1external judge for our decoder

How to read this atlas

A possession model outputs one of three states for every frame: home, away, dead ball. Almost no public dataset contains that answer. Most contain something next to it. Each family below says how close its datasets get, and what they are good for at that distance.

Rule Why it holds
Labels beat pixels A dataset with a possession field teaches more than one that only looks like the target
Pair only what is truly paired Video and tracking belong together when they share a match and a clock
A licence has layers Code, annotations and footage are licensed separately

Our Initial Challenge decoder has no trained weights, so nothing here was a training set. One dataset still decided something: open DFL labels judged the decoder’s parameter profiles and picked the one we submitted.

The families

  1. Diagram: Tracking evidence, Possession label, Home, Away, Dead ball.

    01 of 06

    Frame-level possession labels

    The scarce resource. Three releases put a possession field on the tracking itself, and only one labels all three states under an open licence.

    • The only family that can score a possession model
    • Three providers, three definitions of possession
    • Open DFL labels judged our decoder

    3 entries

  2. Diagram: Player track, Trail.

    02 of 06

    Tracking and full-pitch video

    Positions without possession. They teach what a tactical camera sees and how far a position can be trusted.

    • Position quality is the ceiling on possession quality
    • A track ID is not a player, and a swap looks like a turnover
    • Full-pitch footage under CC BY 4.0 exists

    4 entries

  3. Diagram: Player, Ball.

    03 of 06

    Broadcast video and game state

    Television pictures turned into players, teams and touches. What a possession model consumes is what these benchmarks ask vision to produce.

    • The upstream chain: detect, track, calibrate, identify
    • 102,992 ball actions tied to the player who made them
    • Part of the pitch, and a clock that cuts away

    4 entries

  4. Diagram: Pass, Duel, Restart.

    04 of 06

    Events

    Instants, not intervals. Event data cannot label a frame, but it knows the grammar of possession.

    • Base rates: how long a spell lasts, how many a match holds
    • Restarts mark the edges of the dead ball
    • The acting team is not always the owner

    4 entries

  5. Diagram: Camera view, Calibration, Pitch plane.

    05 of 06

    Detection, calibration and synthetic

    Pixels to metres, and the ball in three dimensions. Possession is decided by distances, so a geometry error is a possession error.

    • Every distance passes through a camera model first
    • Ball height is the dimension 2D possession ignores
    • Synthetic data is for exact geometry and robustness

    5 entries

  6. Diagram: First half, Second half, Broadcast cuts, Own clock.

    06 of 06

    The competition data

    The exam and its predecessor, described as the organisers' public repository documents them. The match files are never published here.

    • 199,319 frames, a ball untracked on 41% of graded frames
    • Players carry a team, not an identity
    • Five public minutes: a check, not a training set

    2 entries