Geometric algebra-based multiview interaction networks for 3D human motion prediction

Abstract

3D skeleton-based human motion prediction is an essential and challenging task for human-machine interactions, which aims to forecasts future poses given a history of their previous motions. Recent works based on Graph Neural Networks (GCNs) show promising performance for motion prediction due to the powerful ability of feature aggregation of GCNs. However, with the deep and multi-stage GCN model deployment, its feature extraction mechanism tends to result in feature similarity over all joints, and degrade the prediction performance. In addition, such a graph structure in recent works was still insufficient to process the high dimensional structural data in Euclidean space when inference through multi-layer networks. To solve the problem, we propose a novel Geometric Algebra-based Multi-view Interaction network (GA-MIN), which captures and aggregates motion features from two interactions: 1) global-interaction, which refactors various spectrum dependencies using geometric algebra-based structure, and 2) self-interaction, which leverage self-attention mechanism to capture compact representations. Extensive experiments are conducted on three public datasets: Human3.6M, CMU Mocap, and 3DPW, which prove that the proposed GA-MIN outperforms state-of-the-art methods on 3D Mean Per Joint Position Error (MPJPE) and Mean Angle Error (MAE) on average.