A multi-object tracking dataset for multiple sports and a study of cross-domain generality

Abstract

Tracking devices that can track players and balls are critical to the performance of sports teams. This paper presents a comprehensive multi-object tracking dataset for sports analysis. Our dataset comprises over 150 minutes of video footage from three sports: football, basketball, and handball. We captured complete pitch view footage using fisheye and drone cameras to enable holistic trajectory analysis. Additionally, we conduct ablation studies on trajectory prediction tasks to investigate the generalizability of learned features across different sports. Our dataset and findings provide valuable insights for sports analysis and can aid in developing advanced tracking and trajectory prediction algorithms.