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

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