03Detection, calibration and synthetic

SoccerNet Camera Calibration

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
Pitch markings in the image, the same markings on the plan. The matrix between them turns pixels into metres.
Modality
Broadcast images with pitch-line and keypoint annotations
Task
Estimate the camera that maps the pitch to the image
Score
Reprojection accuracy within a pixel tolerance, and completeness
Possession labels
None
Video
Stills and clips, not continuous matches
Licence
SoccerNet research terms. Underlying video protected
Access
Annotations and development kit public (SN-Calibration-2023)

What it is

Each image is annotated with the pitch markings it shows: touchlines, penalty areas, the centre circle, the goals. A calibration method has to recover the camera from those correspondences, and is scored on how well the pitch model reprojects onto the image.

How it scores

The score has two parts. Accuracy asks how well the projected markings fall on the annotated ones, within a pixel tolerance. Completeness asks how much of the test set the method answered at all. Scoring both stops a method from looking good by answering only the easy frames, and it measures the failure that matters in production: frames with no answer.

The evaluator is public, so the same test can be run on frames annotated from your own camera. PnLCalib, TVCalib and KpSFR are comparable because they report here.

What it teaches

Signal What you get
Possession Nothing direct
Tracking Nothing
Events Nothing
Calibration The task itself: field markings to a camera model, with a reprojection score

Where it wins

  • Direct supervision. Line and keypoint labels are exactly what a homography estimator consumes.
  • Real broadcast framing. Tight shots, few visible lines, players standing on the markings.
  • A hard score. Methods are judged by reprojecting the pitch model, so a plausible-looking but wrong camera is exposed.

What it costs

  • Single frames. No temporal consistency is tested. A calibration that jitters from frame to frame can still score well.
  • Planar assumptions. One homography models a flat pitch through one lens. Stitched panoramas and lens distortion break it.
  • Occlusion. The fewer markings in view, the less constrained the answer.
  • Pixels, not metres. Reprojection error in the image is not position error on the pitch. A good score still allows large errors far from the camera, and possession cares about metres near the ball.

How it plugs into possession

Possession evidence is measured in metres. Who is nearest the ball, whether anyone is within playing distance, whether the ball has crossed the line. All of it passes through the pixel-to-pitch mapping first. A homography error is a distance error, and a distance error near a touchline is the difference between live and dead.

That makes calibration validity a possession feature in its own right. In the competition data the organisers’ documentation marks the broadcast homography as usable on 65.8% of its frames. A system that uses the mapping only where it can be trusted needs a model of when that is, and this benchmark is where such a model is built.

Our read

Calibration is the quiet dependency of every distance in football AI. This is where it is trained and, more usefully, where its failures are measured.

Sources

  1. Development kit with baseline and evaluatorgithub.com
  2. SoccerNet camera calibration task pagesoccer-net.org
  3. SN-Calibration-2023 datasethuggingface.co
  4. SoccerNet datasets and access tablegithub.com