04Solutions family
Events and replay
Spotters that timestamp passes, drives and tackles, and the tools that untangle a broadcast edit.
- 7entries
- 1 sthe usual scoring tolerance
- 25frames inside that tolerance
One picture: anchors with error bars
An event spotter watches video and emits instants: a pass here, a tackle there. Every change of possession is an event, so spotters are natural possession evidence. Two properties decide how they can be used. They are scored at a tolerance of one second, which is 25 frames at 25 fps. And they are silent between events.
The other half of the family handles the broadcast itself. A broadcast is an edit, with cuts and replays, and a frame inside a replay is not match time.
In the chain, spotting sits beside measuring. Both produce evidence: measuring is dense and geometric, spotting is sparse and semantic.
How to choose inside it
| Entry | Labels | What it needs | Compute |
|---|---|---|---|
| T-DEED, team-aware | Checkpoint available | Broadcast video | GPU |
| 2023 winning solution | Trained models available | Broadcast video. No team in the output | GPU |
| E2E-Spot | Football labels to train | Video | Light GPU |
| FOOTPASS baselines | Labels to train | Video plus identified player tracks | Heavy |
| SoccerNet spotting kits | Not a model | Timestamped predictions to score | Evaluation only |
| PySceneDetect | None | Any video | CPU |
| Replay grounding | Labels to train | Replay segments found upstream | GPU |
For possession, team attribution is the property to pay for. An event with a team is direct evidence. An event without one needs a second source to say whose it was.
What this family cannot do
- It cannot place a boundary to the frame. One second of tolerance is 25 frames of doubt.
- It cannot label the time between events. That is most of the match.
- It cannot be read off a broadcast without hygiene. Cuts and replays must be handled first, or the same tackle is counted three times.
Where our decoder stands
Its tag evidence is used the way this family should be used. The automatic tags are anchors with error bars: restart windows, not instants, and calibrated team signals from every tag code. The decode fills the frames in between and decides where inside each window the boundary falls.
In this family7
-
An event spotter built for temporal precision, extended to say which team acted. The video model whose output is closest to possession evidence.
- Video
- Event spotting
- Open weights
- GPL-3.0
-
The first-place entry of SoccerNet's 2023 ball-action challenge: frames through a 2D encoder, then 3D convolutions, to spot passes and drives.
- Video
- Event spotting
- Open weights
- MIT
-
End-to-end spotting of temporally precise events: per-frame features and a recurrent layer, trained together so that fine timing survives.
- Video
- Event spotting
- Open weights
- BSD-3-Clause
-
Player-centric ball-action spotting: not only what happened and when, but which player did it, using tracks, a graph over players and a denoising sequence model.
- Video + tracks
- Player-level events
- Identity-dependent
- Apache-2.0 code
-
The task definitions, data loaders, baselines and evaluators for action spotting and ball-action spotting. The reference frame every football spotter is measured in.
- Video
- Benchmark
- Evaluation code
-
Shot-boundary detection without learning. Frame-to-frame content change, thresholded, gives the cuts that divide a broadcast into continuous camera takes.
- Video
- Shot detection
- No training
- BSD-3-Clause
-
Given a replay, find the live moment it shows. The piece that stops a broadcast pipeline from counting the same goal twice.
- Video
- Replay
- Benchmark