A Novel Graph Augmentation for Semi-Supervised Learning on Skeleton-Based Action Recognition

In recent years, skeleton-based action recognition has become popular due to its higher data precision and resistance to environmental noise compared to RGB-based methods. However, the cost of annotating accurate skeleton video data is quite high. This difficulty often leads to suboptimal training in supervised learning algorithms. In this study, we propose an efficient data augmentation method for skeleton data and combine it with semi-supervised learning strategies to confront the issue of low-labeled skeleton data in action recognition. Moreover, we optimize the schedule for incorporating unlabeled data into the model training process. This approach ensures both model performance and reduces training time effectively. In experiments utilizing a low amount of labeled data from the NTU RGB+D dataset, our proposed method achieved an accuracy of 81.57% with only 10 labeled samples per class. This compares favorably to the original method's accuracy of 77.5%. Furthermore, with 600 labeled samples, our method improved the original result of 90.31% to 92.03%. Although the gap in improvement narrows as the number of labeled samples increases, our proposed augmentation method successfully enhances the performance of the semi-supervised learning method in skeleton-based action recognition. This opens the door to more efficient labeling in skeleton-based action recognition and improves the algorithm's practicality in real-world scenarios.