Automatic Sport Analysis System for Table-Tennis using Image Recognition Methods

Abstract: This study presents several machine learning models tasked with gathering information about what happens in a table tennis match. This information includes the ball's position, identifying serves, and other events such as ball hits and bounces. The TTNet backbone demonstrated superior performance across all tasks, including inference time. The Classification Head achieved high accuracy (94.6%), F1 score (97.2%), and RMSE value (0.014) for ball detection, accurately predicting ball positions. The TTNet model with the Simple Event Head showed promise for event detection (97.5% accuracy, 80.4% F1 score). However, the dataset is heavily imbalanced and the models would probably benefit from a more balanced training set. Serve detection performance was subpar for all models (best F1 score: 42.1%), requiring significant dataset and model reconstruction. All top-performing models achieved fast inference speeds (<4 ms) on various devices, enabling real-time operation even on smartphones.