Parallel multi-stage rectification networks for 3D skeleton-based motion prediction

It is noted that Recurrent Neural Networks (RNNs), which are widely used in human prediction tasks, have achieved promising performance in motion prediction, owing to RNNs’ robust capacity for spatial-temporal sequence modeling. However, RNN-based methods suffer from error accumulation due to their step-by-step prediction mechanism. Therefore, in this paper, we propose a three-stage parallel prediction network, which guides the output generation of these three networks with different objectives. In particular, we leverage the high-dimensional information in these three networks to fuse new information to generate the final output. In addition, we also designed a fusion block based on GRU and attention mechanism to extract high-dimensional information more efficiently. Extensive experiments show that our approach outperforms most recent methods in both short and long-term motion predictions on Human 3.6M, CMU Mocap, and 3DPW.