02The data

Two cameras, two clocks

The broadcast feed sees the foot on the ball. It also cuts, zooms and replays. Before it can add evidence, every one of its frames needs an honest answer to one question: when was this?

  • Tactical clock
  • Broadcast cuts
  • Replays
Two clocks, one match. A live shot can be slid into alignment; a replay cannot, because it points backwards.
Master clock
Tactical video frame, from 0
Broadcast records
164,488 on their own index
Broadcast ball
Tracked on 50.5% of frames
Mapping
None supplied

One clock rules

The answer is graded on the tactical video’s frame index, starting at 0. Everything on the tactical side hangs off it with simple conventions.

Source Index To tactical frame
Tactical tracking Frame id from 1 Subtract 1
Tactical homography Frame from 0 As is
Smart Tagging Seconds from clip start Multiply by 25 and round
Broadcast, all files Its own frame index No mapping supplied

What the broadcast adds

A director follows the ball. The broadcast view is closer, so feet, contacts and the ball itself are larger in the image, and its tracking is a second, independent look at the same play. The ball is tracked on 50.5% of broadcast frames and its homography is usable on 65.8%.

The price is the edit: pans and zooms, cuts between cameras, close-ups with no pitch lines, graphics, and replays.

Frame numbers do not travel

A broadcast frame cannot be equated with the tactical frame of the same number. Each kind of footage relates to match time differently.

Footage Relation to match time Usable for geometry
Live wide shot Runs forward at clock rate Yes, where the homography is valid
Live close-up Live, but few landmarks Rarely
Replay Jumps back to an earlier event Only once grounded to that event
Slow-motion replay Earlier event, stretched Only once grounded and rescaled
Crowd, bench, graphics No pitch correspondence No

A replay is the dangerous one. It shows a clean view of a turnover that has already happened, on frames that sit next to live play.

How it plugs into possession

Our decoder runs on the tactical clock alone, which keeps every frame of evidence unambiguous in time. Verified broadcast observations fit the same design without changing it: each becomes one more piece of evidence on a tactical frame, weighted by the confidence of its alignment.

Getting there is a design problem of its own: a mapping for each live segment, a confidence on every link, and replays grounded to the passage they repeat or left out. The design is in One clock for the match. The tools are in the Synchronisation family.

Our read

Cross-view fusion is a synchronisation problem before it is a vision problem. Alignment has to be segment by segment, with confidence, and with replays treated as what they are: old news.

Sources

  1. DATA (organisers' repository)github.com
  2. FAQ (organisers' repository)github.com