05The hard parts

Players without names

Tracking slots are not players. A slot can change bodies mid-move and no football has happened. Design around identity, or repair it. Never trust it by default.

  • Tracking slots
  • Set features
  • Re-identification
Two team-mates cross and leave with each other's numbers. A feature read by slot jumps at the swap. The same feature read by team does not move.
Given
A team for every player, no stable identity
Invariant route
Features that ignore slot order
Repair route
Link tracklets by motion and appearance
Trap
Reading a slot change as an event

The design question

What should a possession model do when the data cannot say who a player is? Automatic tracking assigns slots. A slot is neither a stable physical identity nor a shirt number, and an apparent change of slot is not a football event.

The options

Route Evidence it uses Wins Costs
Identity-invariant Nearest-player distances per team, sorted speeds, occupancy near the ball, team shape Unaffected by renumbering No player-level story
Set and graph encoders Deep Sets, Set Transformer over unordered players Learns interactions, order-free Needs labelled matches
Identity repair Motion continuity, feasible speed, Kalman prediction, one-to-one assignment Enables player-level paths Errors compound through a match
Appearance re-identification Visual embeddings from video Recovers after long occlusion Needs video; similar kits confuse it
Role priors Formation slots as soft hints Cheap, tactical Roles are not people

Roles do not rescue identity. Full-backs invert, midfielders interchange, wingers swap flanks, and every set piece dissolves the shape. Shirt numbers live in the pixels, not in the coordinates, and bringing them across is one of the gaps the single source of truth is built to fill.

Repair, or never need it

One player through one crowded box. The tracker delivers five fragments under five numbers. The repair route has to make four links, and one wrong link turns every later frame into another player. The invariant route never needed the links.

What else breaks it in real football

  • A substitution. A new body takes over a slot with a different job.
  • A slot change read as a pass. A player-level model sees the ball “move” between two identities that are the same person.

How to judge it

Run the permutation test: shuffle the slot labels within each team, on every frame, and re-run. An identity-invariant design returns the same output. For repair, fragment clean tracks on purpose and measure how possession accuracy degrades as the fragments get shorter.

Our read

Team possession is a team-level question. Answer it with identity-free evidence, and spend repair effort only when a player-level model needs it.

Our call

ConstraintData reality

Why we reason about teams

Players arrive with a team and no identity. Possession is scored per team, so every evidence feature is identity-invariant: which side has a player within reach of the ball, how still the players are. Nothing in the pipeline depends on a slot staying the same person, and renumbering the slots cannot change the output.

Sources

  1. Zaheer et al. (2017), Deep Setsarxiv.org
  2. Lee et al. (2019), Set Transformerarxiv.org
  3. PathCRF: ball-free possession paths from player trajectoriesgithub.com
  4. SoccerNet Re-Identificationgithub.com