A Transformer-based Approach for Dynamic Referee Assistance

Deep learning-based algorithms have shown great potential in video analysis tasks such as action recognition and event detection. The detection of football events in videos is a growing field that has many benefits, including enhancing the viewing experience for fans and enabling teams to analyze player behavior. In this paper, we present a novel deep learning-based algorithm for accurately and efficiently identifying football events in match videos. Our approach utilizes a combination of NetVLAD++ pooling, early feature-level fused transformer features, and a dynamic hybrid loss function. To evaluate the proposed framework, we create a dataset of 200 football games covering 17 events for training and testing. Through empirical evaluation, we demonstrate the effectiveness of our approach in spotting football events. Our proposed framework has achieved an average mAP of 14.16%, outperforming baseline methods by 11.26%. This breakthrough has significant implications for the ongoing Action Spotting 2023 Challenge and has the potential to revolutionize how football events are detected, analyzed, and acted upon, providing valuable insights to coaches and players.