A Cost-effective Training Scheme for Person Re-Identification in Soccer Matches

In soccer matches, Person Re-identification (ReID) technology plays a crucial role in analyzing the position information and statistics of each player. Besides, multiple cameras are used to capture players during soccer matches. However, the appearance features of the same player in different cameras may exhibit significant differences, which makes person identification difficult. Therefore, a robust ReID model is required to obtain distinctive appearance features and identify each player. Although existing ReID technologies have achieved excellent performance in person identification, they typically require expensive computational resources to train a high-accuracy model for feature extraction. Compared to conventional ReID approaches, this research is built upon the Vision Transformer (ViT), focusing on batch-wise sample imbalance issues to enhance ReID performance within limited batch memory and less training data. Our model training process is performed on a single 24GB Nvidia RTX-3090 graphic card with 23% SoccerNet train data. The experimental results show our approach achieves a ReID performance of 82% in mean average precision (mAP) and 76% in rank1.