Skeleton Graph Scattering Networks for 3D Skeleton-Based Human Motion Prediction
To achieve 3D skeleton-based human motion prediction, many graph-convolution-based methods are proposed for promising results; however, due to only preserving low-pass information over graphs, those graph convolution methods suffer from over-smoothing, causing the predicted poses staying the same in the long term. To resolve the over-smoothing issue, we propose a novel skeleton graph scattering network (SGSN), which leverages graph scattering to extract comprehensive motion information from multiple graph spectrum bands. The core of the proposed SGSN is the adaptive graph scattering block (AGSB), including two key modules: i) graph scattering decomposition, which decomposes information into various graph spectrum bands and updates the trainable features in each band, as well as ii) graph spectrum attention, which aggregates those features in various graph spectrum bands via trainable attention weights. Extensive experiments reveal that SGSN outperforms state-of-the-art methods by 8.5%, 9.0% and 3.9% of 3D mean per joint position error (MPJPE) in average on Human3.6M, CMU Mocap and 3DPW datasets, respectively. We also test the mean angle error (MAE) on Human3.6M, which is lower by 3.3% than previous methods. Moreover, SGSN outperforms even more in the long-term prediction because of the alleviation of the over-smoothing.