02The data

What each data asset contributes

Two cameras, one clock that counts. What every provided asset can tell a possession model, and where each one stops.

  • Tactical
  • Broadcast
  • Smart Tagging
  • Homography
The tactical layers share one frame index. The broadcast layers sit on their own clock until someone proves the alignment.
Reference clock
Tactical video, 25 fps
Geometry
Player and ball tracking at 25 Hz
Context
638 automatic tag intervals
Cross-camera sync
Not provided

The inventory

Asset What it contributes Where it stops
Tactical video The reference: 199,319 frames, whole pitch The ball is tiny in the image
Tactical tracking 25 Hz positions and speeds for players and ball Ball missing, slots unstable
Tactical homography Pitch-to-image transform for every frame Planar: no height for an airborne ball
Smart Tagging 638 automatic tactical intervals and restarts Phases, not a pass-by-pass event log
Broadcast video Close views of feet, contacts and the ball Cuts, close-ups and replays
Broadcast tracking A second numerical view of players and ball Its own slots, its own clock
Broadcast homography Ground projection for the broadcast view Usable on 65.8% of frames

Each camera’s video, tracking and homography are aligned to that camera’s own frame index. No correspondence between the two cameras is provided.

What carries the signal

Possession is geometry over time, so the tactical tracking is the principal source: who is near the ball, how fast the ball is moving, whether it has left the pitch, whether the players have stopped. Its weakness defines the challenge. The ball is untracked on 41% of graded frames, and a player is a team and a slot, not an identity. See tactical tracking and the ball goes missing.

Smart Tagging answers a different question. It marks set pieces and restarts, which is direct evidence about dead ball, and it names a team for tactical phases, which is weak evidence about who has it. See tags are context, not events.

The homography is the bridge between pitch coordinates and pixels. It is what lets anything seen in the image be placed on the pitch. See homography.

More is not automatically better

The broadcast feed sees what the tactical camera cannot: the foot on the ball. It also runs on an edited timeline. A broadcast observation joined to the wrong tactical frame is not a noisy measurement; it is a confident false one, and at 25 frames per second a small offset is enough to invent a turnover.

A unified timeline has to carry its uncertainty: which cross-view matches are verified, which are plausible, which are absent. See two cameras, two clocks.

What fed our decoder

The evidence decoder reads the tactical tracking and the Smart Tagging intervals, both on the tactical clock. Nothing else is needed to produce the submission, which is why a full match decodes in under 10 seconds on a laptop CPU.

Open data from outside the competition did a different job: choosing parameters. That atlas lives in Datasets.

Our read

The tactical tracking carries the geometry, the tags carry the restarts, and everything else is corroboration that has to be earned. More sources only help when they are provably about the same moment.

Sources

  1. DATA (organisers' repository)github.com