Abstract
In human skeleton-based action recognition, graph convolutional networks (GCN) have shown significant success. However, existing state-of-the-art methods struggle with complex actions, such as figure skating, where performance is often unsatisfactory. This issue arises from two main factors: the lack of shift, scale, and rotation invariance in GCN, making them especially vulnerable to perspective distortions in 2D coordinates, and the high variability in displacement velocity, which depends more on the athlete's individual capabilities than the actions themselves, reducing the effectiveness of motion information. To address these challenges, we propose a novel cosine stream to enhance the robustness of spatial features and introduce a Keyframe Sampling algorithm for more effective temporal feature extraction, eliminating the need for motion information. Our methods do not require modifications to the backbone. Experiments on the FSD-10, FineGYM, and NTU RGB+D datasets demonstrate a 2.6% improvement in Top-1 accuracy on the FSD-10 figure skating dataset compared to current state-of-the-art methods. The code has been made available at: https://github.com/Jiahao-Guan/pyskl_cosine.