Camera Calibration and Player Localization in SoccerNet-v2 and Investigation of their Representations for Action Spotting

Abstract

Soccer broadcast video understanding has been drawing a lot of attention in recent years within data scientists and industrial companies. This is mainly due to the lucrative potential unlocked by effective deep learning techniques developed in the field of computer vision. In this work, we focus on the topic of camera calibration and on its current limitations for the scientific community. More precisely, we tackle the absence of a large-scale calibration dataset and of a public calibration network trained on such a dataset. Specifically, we distill a powerful commercial calibration tool in a recent neural network architecture on the largescale SoccerNet dataset, composed of untrimmed broadcast videos of 500 soccer games. We further release our distilled network, and leverage it to provide 3 ways of representing the calibration results along with player localization. Finally, we exploit those representations within the current best architecture for the action spotting task of SoccerNet-v2, and achieve new state-of-the-art performances.

Dynamic Multi-Person Mesh Recovery From Uncalibrated Multi-View Cameras

“…However, the dynamic human bodies occupy the most proportion of the image pixels in multiperson scenarios. To handle this obstacle, [50,12,8,50,13] obtain structure cues and estimate camera parameters from the semantics of the scene (e.g., lines of the basketball court). [24,55] estimate the extrinsic camera parameters from the tracked human trajectories in more general multiperson scenes.…”

“…We then qualitatively and quantitatively evaluate the estimated camera parameters. Since there exists a rigid transformation between the predicted camera parameters and the ground-truth provided in the datasets, we follow [12] to apply rigid alignment to the estimated cameras. We first com- ) since it relies on the dense correspondences between each view.…”