03Frame-level possession

SkillCorner Open Data

Ten A-League matches of broadcast-derived tracking with a possession field, phases of play and, crucially, a flag that says which positions were seen and which were extrapolated.

  • Tracking
  • Possession field
  • Detected or extrapolated
Inside the camera view a position is detected. Outside it, the position is extrapolated, and the data says so.
Modality
About 10 Hz broadcast-derived player and ball tracking
Scale
10 A-League Men matches, 2024/25
Extras
Dynamic events, phases of play, season aggregates, two body-pose examples
Possession labels
Per-frame possession field. No explicit dead-ball state
Quality flags
is_detected separates observed from extrapolated positions
Video
None
Calibration
Camera-view polygon per frame, no full homography
Licence
MIT on the repository
Access
Public GitHub repository

What it is

SkillCorner tracks football from the broadcast feed, and this repository opens a sample of that product: ten A-League Men matches with player and ball positions at about 10 Hz, team and player identifiers, a possession entry per frame, derived dynamic events and phases of play.

The detail that matters is is_detected. Broadcast cameras show part of the pitch. SkillCorner extrapolates the rest and says so, per object, per frame.

What it teaches

Signal What you get
Possession A per-frame possession field, plus phases of play as side supervision
Tracking Broadcast-domain tracks with observed and extrapolated positions kept apart
Events Derived dynamic events aligned to the tracking
Calibration The visible pitch area per frame, as a polygon

Where it wins

  • It models its own uncertainty. Observed and inferred positions are separated in the data, not blended.
  • It behaves like automatic tracking. Gaps, partial views and filled-in players are the normal case, as they are in any feed that was not captured by in-stadium hardware.
  • Phases come free. Build-up, transition and the rest arrive as labels, useful context for when possession is likely to change.

What it costs

  • No dead-ball label. The possession field says who, not whether the ball is alive. A gap between phases is not a stoppage, and treating it as one manufactures dead-ball frames.
  • 10 Hz, not 25. Short touches and quick turnovers sit between samples. Upsampling invents precision that was never measured.
  • Ten matches, one league. A test bench, not a training corpus.

How it plugs into possession

Use it to answer one question: what should a model do when the evidence disappears? Automatic tracking loses the ball constantly. In the competition match it is untracked on 41% of graded frames. A decoder has to read a missing ball as no evidence, not as a stoppage.

SkillCorner is the open dataset where that behaviour can be studied with the truth attached: mask on the detection flag, compare possession calls on detected frames against extrapolated ones, and see how far a model’s confidence should fall when it is working from inference.

Our read

The most honest tracking in the open. It tells you what the camera saw and what the model filled in, which is exactly the distinction a possession decoder needs when the ball goes missing.

Sources

  1. Official repository and schemagithub.com
  2. Repository licencegithub.com
  3. Match indexgithub.com