03Datasets family
Detection, calibration and synthetic
Pixels to metres, and the ball in three dimensions. Possession is decided by distances, so a geometry error is a possession error waiting to happen.
- 5datasets
- 1.5 mcontrol radius our decoder keeps physical
- 2sources of real 3D ball positions
What this family is
Five datasets with no football semantics at all. They do not know what a pass is. They know where the pitch is in the image, where a player stands in metres, how high the ball is, and how to keep a detector working in bad light.
That is exactly why they matter to possession. Control is a distance. Out of play is a line. Both are measured after a camera model has turned pixels into metres.
Five datasets, four jobs
| Dataset | Job | Footage | Licence |
|---|---|---|---|
| SoccerNet Calibration | Camera from pitch markings | Real broadcast | Research terms |
| SoccerNet-v3D | 3D ball from a live shot and its replay | Real broadcast | Footage on SoccerNet terms |
| ball3d | 3D ball trajectories from many views | Real, one match at its core | CC BY-NC-SA 4.0 |
| Spiideo SynLoc | Player positions in metres from one frame | Synthetic and real | Set by Spiideo |
| SoccerSynth-Detection | Player detection under varied conditions | Synthetic | Zenodo record terms |
How to choose
| You need | Use |
|---|---|
| A homography estimator, and a measure of when it fails | SoccerNet Calibration |
| Ball height from ordinary television pictures | SoccerNet-v3D |
| A ball motion model tested against truth | ball3d |
| A detector scored in metres, not pixels | Spiideo SynLoc |
| Robustness to light, texture and blur | SoccerSynth-Detection |
Height and metres
Two quantities decide more possession calls than any model choice. The first is ball height. In 2D a long ball passes within reach of several players on its way down the pitch, and none of them can play it.
The second is localisation error. A box two pixels off at the far touchline of a wide view is a metre off on the grass.
The traps
- Matrix direction. ball3d’s homographies map image to pitch. Other providers ship pitch to image. Check before composing transforms.
- One plane, one lens. A homography assumes a flat pitch through a single camera. Stitched panoramas and lens distortion break it.
- Single frames. Calibration benchmarks score each image alone. A camera model that jitters between frames can still score well.
- Sim to real. Rendered players are cleaner than real ones. Synthetic data widens a detector; it does not validate one.
- Licences. ball3d is non-commercial and share-alike. The synthetic sets carry their hosts’ terms.
What it cannot do
Tell you anything about football. No events, no teams in possession, no spells. This family improves the evidence. It never supplies the answer.
Where it meets our decoder
Our decoder works in pitch coordinates and keeps its control radius physical: a player within 1.5 m of the ball and closer than any opponent. The radius stayed outside the parameter search on purpose. Left free, the search on simulated matches pushed it to an unrealistic 0.8 m.
A missing ball is treated as no evidence, never as a stoppage, which is the defensive version of this family’s lesson: when the geometry cannot say, do not guess. What calibration has to guarantee before two cameras can be used as one is set out in Space is not time.
In this family5
-
Broadcast images annotated with pitch lines and keypoints, and the benchmark every football calibrator reports on. The dataset behind most homography models, and the exam they sit.
- Broadcast images
- Pitch lines
- Camera model
-
A benchmark that treats a broadcast replay as a second camera: calibrated views of the same moment, and the 3D ball position they pin down.
- Broadcast and replays
- Camera geometry
- 3D ball
-
A physics-based 3D ball benchmark: a full professional match seen by one broadcast and six static cameras, with triangulated ball trajectories, calibration and ball-out signals.
- Multi-view video
- 3D ball
- CC BY-NC-SA 4.0
-
A localisation benchmark built from synthetic images with exact answers: find every player in a wide-angle frame and say where they stand on the pitch, in metres.
- Synthetic images
- World coordinates
- Single frame
-
A synthetic, broadcast-style image corpus for player detection, rendered with controlled lighting, textures, camera angles and motion blur.
- Synthetic images
- Player boxes
- Zenodo