ClassRAC: Leveraging Action Classification to Improve Repetitive Action Counting Accuracy

The counting of repetitive actions is widely used in physical exercise for sports training and assessment. Existing methods have mainly concentrated on training models across multiple action classes, often neglecting the identification of specific action types. Building on previous experiments, we discovered that separating the training and validation datasets by action classes and conducting individual training and validation for each class can significantly enhance model performance. Inspired by this finding, we divide the task of repetitive action counting into two components: action classification and the optimization of counting algorithms for each action class. Specifically, we develop an action classification algorithm based on prior work, achieving an accuracy of 99.3% on the RepCount dataset. We also optimize the counting method for each action class, resulting in a Mean Absolute Error (MAE) of 0.146 and an Off-By-One (OBO) accuracy of 0.781 on the same dataset, which substantially surpasses the current state-of-the-art (SOTA) methods. Additionally, we introduce two preprocessing techniques tailored for action videos. Experiments confirm that these preprocessing methods effectively filter out irrelevant information, thereby improving counting accuracy.