Offensive Play Recognition of Basketball Video Footage Using ActionFormer

This chapter delves into the use of advanced computer vision techniques to recognize offensive plays in basketball video footage. It begins with an introduction to the role of technology in sports and the importance of coaching. The background section explains the phases of basketball plays and the complexity of action recognition. The study justifies the need for content-based video indexing and the benefits it brings to various stakeholders in basketball. The dataset section discusses popular action recognition datasets and highlights the SpaceJam dataset as the most suitable for this study. Similar work in the field is reviewed, including recent developments in action recognition using transformers and 3D-CNNs. The experiment setup details the pre-processing steps and the models used, including Temporal Segment Networks, Inflated-3D, and Pose Estimated Convolutional 3D. The results section compares the performance of these models on the SpaceJam dataset, showing promising accuracy in distinguishing between actions. The conclusion summarizes the findings and the potential of these models in the context of basketball, emphasizing their ability to be fine-tuned for specific domains.