Soccer Pitch Camera Calibration via Scene Coordinate Regression

Abstract

Many sports broadcasting technologies rely on calibrated cameras to provide accurate spatial information, enabling the analysis of actions during a sporting event. This thesis was done in collaboration with EVS, a company that develops sports broadcasting equipment. Calibrated cameras are useful for EVS VAR solutions, notably for drawing the offside line in soccer matches, and could potentially be used in the future to review actions in 3D using technologies such as Gaussian splatting. These applications highlight the critical role of camera calibration.

This thesis investigates the problem of calibrating soccer pitch cameras using scene coordinate regression (SCR) models, which allow camera parameters estimation from a single image, when combined with PnP+RANSAC. This approach is motivated by the expectation that this operation can be performed quickly. The work focuses in particular on ACE and GLACE SCR models. After a review of camera modeling and calibration methods, as well as a presentation of the main challenges posed by soccer pitch scenes in this context, the work evaluates the iterative ACE0 algorithm and compares it to state of the art structure from motion (SfM) methods, which are found to struggle to provide high-quality camera parameters for reliable application.

The thesis then shifts focus from the ACE0 SfM algorithm to SCR models trained on groundtruth camera parameters, demonstrating that ACE achieves significantly better results, with low angular and position error for most images. The work first tests ACE combined with P3P+RANSAC, achieving very high pose accuracy but without providing the focal length, and then evaluates ACE with P4Pf+RANSAC, which enables focal length estimation but with reduced pose estimation accuracy. Finally, GLACE improves results for challenging camera positions by leveraging global image encodings and cluster-based pose decoders, although this comes with increased computational cost.