Progressive Semi-Supervised Learning for Enhanced Detection in Tennis Sport: From Simple to Complex

Recent progress in the field of sports analytics underscores the critical importance of object detection, with a particular emphasis on the detection of ball sports. However, major state-of-the-art methods based on supervised learning require large annotated datasets, which are often scarce and difficult to obtain. In this paper, we explore an innovative approach to semi-supervised learning in the specific context of tennis sport. This method employs a progressive semi-supervised learning framework, starting from simple individual object detection, such as tennis balls and rackets, gradually transitioning to more complex scenarios involving combinations of rackets and players, and ultimately achieving comprehensive multi-object detection in tennis scenes. This phased training strategy can help the model gradually and efficiently generate high-quality pseudo-labels for large amounts of unlabeled data. Comprehensive experiments on both the MS-COCO dataset and our custom dataset demonstrate the superior performance of our framework. Specifically, our method outperforms the state-of-the-art method by 0.73, 1.45, and 1.3 mAP on the MS-COCO dataset when respectively using 2%, 5%, and 10% labeled data.