Robustness Evaluation of Pre-trained vs. Fine-Tuned Computer Vision Models for Score Detection in Dynamic Sports Environments

This paper explores the efficacy of two distinct approaches in computer vision, particularly in intelligent sports training systems. Our focus is on the development and comparison of two methodologies for image recognition in a basketball score detection system, named “Koško.” The first approach uses a pre-trained model with two combined datasets in the training phase. This method leverages diverse data to enhance the model’s ability to generalize and accurately recognize scores in basketball games. In contrast, the second approach fine-tunes a model initially trained on one dataset with a subsequent second dataset. Rooted in recent transfer learning studies, this method shows that selectively fine-tuning a subset of layers or customizing learning rates can significantly improve robustness to out-of-distribution (OOD) data. By enforcing weight projections, this approach aims to maintain the generalization capabilities of the pre-trained models while adapting to the new dataset. Our research balances precise score detection and pragmatic system development, contributing to intelligent sports training systems. This work highlights the feasibility of combining machine learning and vision-based methodologies to advance sports training systems. Research on basketball score detection using the YOLOv8n model shows the pre-trained model achieved 93.75% accuracy and a 3% false positive rate, outperforming the fine-tuned model’s 88.71% accuracy and 5% false positive rate. By evaluating both networks, this study aims to provide clear strategies for implementing computer vision systems in sports, particularly in the dynamic environment of basketball.