Reciprocal collaboration network for 3D skeleton-based human motion prediction

Abstract

Noticing that the human motion in both the positive and inverted time sequence are predictable, we design a reciprocal network which contains a forward and a backward prediction network to learn this information respectively, and then train the two networks in a collaborative and iterative way with the guidance of consistency constraint. Based on the reciprocal network, we establish an error compensation network to approximate the error of forward prediction network for making the forecast closer to the real pose. Extensive experiments show that our approach outperforms most recent methods in both short and long term motion predictions on Human 3.6M and CMU Mocap.