02The problem

Why possession is a decoding problem

A classifier answers what a frame looks like. A decoder answers which sequence of states best explains the whole half. Possession needs the second.

  • Sequence not frames
  • Offline
  • Three states
Three states per frame, one best path per half. A change of state has a price, so one noisy frame cannot buy a turnover.
States
H, A, D
Scored unit
A whole possession
Mode
Offline: past and future evidence
Short spells
A recovery can last 0.8 s, 20 frames

Classifier, decoder, optimiser

An instantaneous classifier estimates which state looks likely at one moment. A temporal decoder chooses a consistent sequence, weighing evidence across time against the cost of changing state. A temporal optimiser is the general case: any search for the best sequence under a score.

Possession is defined by persistence, so the decoder is not post-processing. It is the model.

The same frame, four answers

A frame carries positions, not a history. Each of these looks the same to a classifier whichever way the match went.

What one frame shows The label depends on
The ball two metres from an opponent Whether it is a pass in flight or a ball he controls
An opponent touching the ball Whether he keeps it or it runs on to the same team
No ball in the tracking What was happening before it vanished
A still ball beside a player Whether play stopped a second ago

None of the four can be settled inside the frame. All four are settled by the frames around it.

Offline changes the problem

A live system filters: it sees the past and the present. The challenge hands over complete recordings, so the decode can smooth: every frame is judged with the frames after it as well. A reception three frames later confirms the pass; a restart two seconds later confirms the stoppage. Decoding a whole half at once uses that for free.

Mid-pass, the nearest player is an opponent. With the past only, a live system leans away. With the reception three frames later in view, an offline decode keeps the passer's team.

The failure that matters

A counter-press recovery can last 0.8 seconds: win it, lose it. To a smoother that is twenty frames of noise. To the scorer it is a possession, and deleting it also merges the two spells on either side into one that matches neither.

A sequence that looks cleaner can score worse. That is why the price of a switch cannot be set by eye.

From the problem to the design

Which decoder to run, and at what price per switch, is a design decision. The candidates are compared in Seven ways to decode a match. The algorithm we chose is explained under Viterbi decoding. The fielded system, with its two switch costs, is The Shrewdies decoder. Duration-aware decoding for the shortest possessions is on the roadmap.

Our read

The risk is not noise, it is over-correction. Every smoother deletes flicker first and real short possessions second. The decoder has to make that trade explicitly, with a price per switch, and the price has to be chosen on real labels.

Sources

  1. Rabiner (1989), A tutorial on hidden Markov modelsdoi.org
  2. EVALUATION (organisers' repository)github.com