A GRAPH-BASED METHOD FOR SOCCER ACTION SPOTTING USING UNSUPERVISED PLAYER CLASSIFICATION.

In this presentation we talk about our graph-based model for action spotting in soccer games. The goal of action spotting is to identify the specific time an action occurs. Our model represents players, referees, and goalkeepers as nodes in a graph and connects according to their distance in the field. We describe in detail our methods for extracting this information from the frames and each component of our architecture. Finally, we present the experiments we carry on the SoccerNetV2 benchmark and show that our performance (57.83% mAP accuracy) surpasses similar graph-based methods and has competitive results with heavy computing methods.