05The hard parts

One clock for the match

Every scored frame is a tactical frame. Any other source must be mapped onto that clock before it can testify, and a replay makes the mapping run backwards. Which evidence can fix the mapping, and what the mapping must record.

  • Synchronisation
  • Broadcast
  • Replays
The tactical clock runs straight. The broadcast is cut into segments: live ones map forward, the replay points back to an instant already played, the crowd shot maps to nothing.
Reference clock
The tactical frame index
Mapping
Per segment, with a type and an uncertainty
Segment types
Live, replay, slow motion, unknown
Trap
One global offset for a whole broadcast

The design question

Which frames from two sources show the same instant? Until that is answered, a second camera is a separate film. What the broadcast feed is, and why it cuts and replays, is described in Two cameras, two clocks. This study is about the design of the mapping: the first prerequisite of the single source of truth.

Evidence for alignment

Evidence Strength Limit
Native timecode Exact when present Often absent or reset at cuts
On-screen match clock Readable anchor Hidden in replays and close-ups
Shared audio Dense, continuous Needs the same audio in both feeds
Movement fingerprints Uses the play itself Ambiguous in static phases
Event anchors with a robust offset fit A few good anchors fix a segment Anchors must be independent
Constrained dynamic time warping Handles drift Invents matches without constraints
Learned embeddings Aligns without hand features Training data, opaque errors

The mapping record

One record per broadcast segment: the tactical frame range it corresponds to, its type, the evidence that produced the link, and the timing uncertainty. Downstream stages read the uncertainty before the frame.

Uncertainty inside a segment

Two anchors lock one live segment. Between them the clocks drift, so the timing uncertainty is smallest at an anchor and widest half-way. Slow motion changes the frame ratio, and no single offset describes it.

What breaks it in real football

  • A replay is not an offset. The mapping is non-monotonic across segments, and one tactical instant can appear on screen more than once.
  • Crowd shots have no match in them. A bench or the stands cannot be aligned to play. Leave them unmapped.

How to judge it

Check alignment against contact anchors that were not used to fit it: a kick-off, a header, a bounce. Report the error in frames, per segment type, and prefer an honest gap to a confident wrong link.

Our read

Synchronisation is a per-segment mapping with a confidence, never a single offset. When the evidence is thin, a candidate interval is the correct answer and a frame number is a false one.

Sources

  1. More Than a Hack 2027: the datagithub.com
  2. dtw-python: dynamic time warpinggithub.com
  3. SoccerNet Replay Groundinggithub.com
  4. Dwibedi et al. (2019), temporal cycle-consistency learningopenaccess.thecvf.com