04Solutions family
Synchronisation
Putting video, audio and tracking on one clock, from file timestamps to geometry.
- 5entries
- 0.4 sboundary shift from a 10-frame offset
- 3need no training
One picture: two clocks, one match
Every source of football data keeps its own time. Tracking counts frames. A video file stores a timestamp for each frame. Events are logged by people. The picture above is the whole problem: two timelines of the same match that must be slid, and sometimes stretched, until they agree.
The price of getting it wrong is concrete. At 25 fps an offset of ten frames moves every possession boundary by 0.4 seconds, and any evidence that combines two sources ends up comparing two different moments.
Syncing comes before any evidence that mixes sources. With a single feed it is already done.
How to choose inside it
| Entry | Labels | What the sources must share | Compute |
|---|---|---|---|
| FFprobe and PyAV | None | Nothing. It reads one file’s clock | Seconds |
| Audio offset estimation | None | Sound from the same stadium | Light |
| Dynamic time warping | None | A signal both measure | Light with a window |
| Geometric synchronisation | None for the principle | Moving players seen by both | Moderate |
| Learned video alignment | Multi-view footage to train on | The same action | GPU |
Work from cheap to expensive, top to bottom. Most problems end at the second or third row.
What this family cannot do
- It cannot find one offset where none exists. An edited broadcast has a different offset after every cut.
- It cannot align what the sources do not share. No common sound, no common view, no common signal: no synchronisation.
- It cannot beat its features. Alignment is only as sharp as the signal being aligned.
Where our decoder stands
On one clock by construction. The task is one label per frame, 199,319 frames at 25 fps, and the decoder labels those frames. Its one time-uncertain input is the tag stream, which it reads as windows, not instants. Synchronisation becomes a design problem the moment video evidence is added.
In this family5
-
Read what the video file says about time: presentation timestamps, time base, real frame rate. The first step of any synchronisation, and the one most often skipped.
- Video
- Timebase
- No training
- LGPL / BSD
-
Two recordings of the same whistle differ by a delay. Generalised cross-correlation finds it far below the length of a video frame.
- Audio
- Alignment
- Classical
- MIT
-
Aligns two sequences that show the same thing at different or varying speeds, by finding the cheapest monotonic path between them.
- Time series
- Alignment
- Classical
- GPL-3.0
-
Time offset as a geometric unknown. If two cameras see the same moving points, the offset that makes their geometry consistent is the right one.
- Video
- Multi-camera
- Geometry
- Research
-
Self-supervised embeddings in which the same moment looks the same from any camera. Alignment becomes nearest-neighbour search in feature space.
- Video
- Cross-view sync
- Learned
- Research code