Goal! A practical guide to soccer video understanding
Anthony Cioppa (ULiège), Silvio Giancola (KAUST), Adrien Deliège (ULiège), Floriane Magera (EVS Broadcast Equipment and ULiège), Vladimir Somers (UCLouvain, EPFL, and SportRadar), Le Kang (Baidu Research), Xin Zhou (Baidu Research), Bernard Ghanem (KAUST), and Marc Van Droogenbroeck (ULiège)
Date: Wednesday, June 1 @ 13:00-16:00
The SoccerNet dataset released in 2018 marked the start of large-scale soccer analysis in academia, gathering a growing research community which now expands to the industry. Broadcast soccer video understanding is an attractive topic for graduate students with many potential applications, like highlights composition and statistics generation. Besides, it encompasses natural yet challenging tasks for computer vision professionals, such as action spotting, camera calibration, player re-identification and tracking. It also comes with specific difficulties to handle fast-paced actions, players of similar appearance and replays through various camera views. All these aspects make soccer a rich yet often overlooked playground for research.
This tutorial focuses on the practical side of building soccer video understanding pipelines: which data is available, how to annotate it, how to use it, which useful tasks can be defined, tackled, and assessed, and which challenges keep the community and industries busy. Demos with Python code will be presented step-by-step to cover a large panel of soccer-related tasks. The instructors and presenters of the tutorial are experienced scientists from academia and industry that lead the soccer research community and develop cutting-edge technologies for sports broadcasts.
This tutorial is tailored for computer vision master students and their professors seeking computer vision classes or thesis projects, for PhD candidates focusing on spatio-temporal aspects of video analysis, for researchers and industrials willing to apply AI techniques within sports broadcasts, and for any soccer enthusiast. The download information of the SoccerNet dataset indicates that all those types of profiles regularly use the dataset. The tutorial assumes basic knowledge of Python and neural networks. Upon completion of the tutorial, attendees will have at hand various pipelines to tackle tasks such as action spotting, player tracking, player re-identification, camera calibration, that they can use not only in soccer-related projects but also transfer to their own research. All the material produced within the tutorial will be made available online.
|13:35||Introducing massive annotations for soccer|
|15:15||Action spotting and replay grounding|
|15:40||Challenges, Q&A and future directions|