Computer Umpires: A Comparative Analysis of YOLOv9 and YOLOv8 for Real-Time Ball Tracking in Table Tennis

Abstract

Real-time precise ball tracking in table tennis is of crucial importance to automatic umpiring and performance analysis, and yet remains technically challenging due to the smallness, fastness and blurriness of the moving ball. Despite improvements in existing systems, a trade-off between high precision and real time application is enigmatic, with professional tournaments still relying on human umpires and a restrictive player challenge mechanism. In this work, we examine the performance of the state-of-the-art YOLOv9 object detector for real-time table tennis ball tracking, in comparison with its predecessor YOLOv8. This work is intended to explore the compromise between precision and recall, and to make a clear recommendation for model choice in different sports analytics. The YOLOv9-C and YOLOv8 models were trained and tested using a publicly available table tennis image dataset. Both the models are trained in the same way (30 epochs; input size 640 × 640), so that the comparison is fair. Standard object detection measurements including Precision, Recall and mAP@0.5). The experiment results show the greatly improved precision (0.880 vs 0.743) and mAP@0.5 (0.613 vs 0.579) with respect to YOLOv8. On the other hand, YOLOv8 had higher recall value of 0.604 compared to 0.508 for YOLOv3, suggesting a higher number of true ball instances could have been found, however also caused a higher number of false positives. Results show that YOLOv9, with higher prediction accuracy than Fast-YOLO, is more suitable for use with low false positive tolerance systems, like automatic umpiring and rule verification. YOLOv8, on the other hand, with its higher recall should be considered for applications such as post-game trajectory analysis, which requires complete event extraction. These findings contribute significantly to a broader utilization of deep learning in high-speed sports and suggest potential future work such as incorporating temporal data to improve tracking robustness.