MHFDN:Multi-branch Hybrid Frequency Domain Networks for 3D Human Motion Prediction

Recent studies have shown great potential in 3D human motion prediction using Graph Convolutional Networks (GCN)-based modeling. However, these methods have two key problems: First, there is a lack of explicit extraction of the spatial features of joint data. Second, the relevant high-frequency components are often ignored in the feature extraction process. In order to solve these two problems, we propose a novel and universal hybrid frequency domain networks based on multi-branch structure, which contains multiple mixed frequency domain blocks to give more focus on high frequencies information during training. Through the multi-branch structure, we can learn both local and global movements of the human body. Since the bottom layer is mainly responsible for capturing high-frequency details, the top layer is more concerned with the modeling of low-frequency global information, we further introduce the structure of frequency component modification module and perform information trade-offs based on different networks layers, enabling each layer to effectively extract appropriate high-frequency and low-frequency human features. Experiments show that our method has high prediction performance and higher robustness on the Human3.6M and CMU Mocap datasets.