Improved Accuracy of Table Tennis Ball Detection by Fine Tuning

In recent years, image recognition and deep learning have begun to be widely introduced into the field of sports competitions. In table tennis, ball detection is crucial for analysis, and more accurate detection is required. The purpose of this paper is to enable table tennis ball detection with higher accuracy using fine tuning. In the performance evaluation, we evaluated object detection accuracy and inference time. We also evaluated retraining using three pre-learned models on the table tennis dataset, and included existing methods to compare their performance. Results of the experiments showed that all of the fine-tuned models using table tennis ball dataset improved over existing table tennis ball detection methods in terms of both accuracy and inference time. In particular, the best was achieved when the pre-trained model used Mask R-CNN that have been trained on large datasets.