03Events

SoccerNet Ball Action Spotting

Dense, timestamped ball actions on broadcast video: passes, drives, headers, crosses, shots, throw-ins, tackles and ball out of play. The atoms a possession is made of.

  • Broadcast video
  • 12 action classes
  • Research terms
Each ball action is a candidate boundary. Possession changes at some of them, never between them.
Modality
Broadcast video with timestamped ball actions
Classes
12 in recent editions, from pass and drive to tackle and ball out of play
Possession labels
None. Single timestamps, no spells
Calibration
None
Licence
SoccerNet research terms, non-commercial. Video under NDA
Access
Labels and development kit public

What it is

A SoccerNet challenge track that moves action spotting down to the ball. Where the original task labels a few sparse match events, this one timestamps the actions that make up open play, and scores models on finding them to within a tight tolerance.

What it teaches

Signal What you get
Possession Candidate change points: every touch that could move the ball between teams
Tracking Nothing
Events Twelve classes of ball action, densely annotated
Calibration Nothing

Where it wins

  • Temporal precision. Tight tolerances force models to localise the touch, not the passage of play.
  • The right classes. Pass, drive, tackle and ball out are exactly the events a possession spell starts or ends on.
  • Video-native. It trains detectors that work from the broadcast picture, with no tracking required.

What it costs

  • No actor. The label says a pass happened, not which player made it. FOOTPASS adds that.
  • Timestamps, not intervals. Unlabelled time is not dead ball and not “no possession”. Filling it with either label manufactures errors.
  • Gated video. The footage sits under SoccerNet’s non-commercial terms.

How it plugs into possession

A possession sequence is piecewise constant. It can only change at a touch or when the ball leaves play. A ball-action detector therefore proposes where the boundaries may be, and a decoder decides which of them are real.

That division of labour is sound whatever the input: let fast, local evidence nominate change points, and let a global decode over the half choose among them. In our decoder the nominations come from tracking geometry and tactical tags; the principle is the same.

Our read

The right granularity for possession. Not goals and cards but passes and touches, the events where the ball actually changes feet.

Sources

  1. Task and coveragesoccer-net.org
  2. SoccerNet repository and access matrixgithub.com
  3. SoccerNet FAQ on rightssoccer-net.org