03Detection, calibration and synthetic

SoccerNet-v3D

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
The live shot and the replays see the same instant from different places. Their rays meet at the ball.
Modality
Broadcast and replay views with camera calibration and 3D scene annotations
Task
Camera geometry and 3D ball reconstruction
Possession labels
None
Video
Still views of single moments, not continuous matches
Licence
Public repository. Footage keeps its original SoccerNet terms
Access
Public GitHub repository with data links

What it is

Broadcasters show important moments twice: live, and again from another angle. SoccerNet-v3D uses that. With both views calibrated against the pitch, the same instant is seen from two positions, and the scene, the ball above all, can be reconstructed in three dimensions.

What it teaches

Signal What you get
Possession Nothing direct
Tracking Ball position in 3D at isolated instants
Events Nothing
Calibration Camera models and reprojection supervision across views

Where it wins

  • Height from broadcast. No stadium rig, no special cameras. The second view comes free with the production.
  • Geometry you can check. Reprojection error across two views is a hard test of a calibration.
  • Monocular supervision. 3D labels to train single-view ball-height estimators against.

What it costs

  • Moments, not sequences. No continuous trajectory, no match clock.
  • Replay bias. Replays show shots, fouls and goals. Ordinary open play is under-represented.
  • Separate rights. The repository is public; the footage keeps its own terms.

How it plugs into possession

Height is the dimension possession models most often ignore. In 2D, a goal kick passes “within 1.5 m” of several players on its way down the pitch. None of them can play it. A model that knows the ball is eight metres up does not hand it to the player underneath.

Tracking feeds rarely carry reliable ball height, so the practical lesson is defensive: treat fast, long, straight ball movement as flight, and do not let proximity during flight change possession. Datasets like this one are where a height estimator that removes the guesswork would be trained.

Our read

A neat idea with a narrow yield. It will not teach possession, but it is one of the few sources of real 3D ball positions from ordinary television pictures.

Sources

  1. Repository and papergithub.com