Component-wise Self-Correction Network for Human Motion Prediction

Abstract

Human motion prediction is a fundamental task in human-robot interaction and self-driving. Many existing human motion prediction methods use one encoder to embed the historical human poses and one decoder to predict future motion poses. We believe that it is possible to estimate the deviation of the decoding results from the groundtruth and use this estimation to further enhance the prediction results. A single deviation estimator cannot deal with all the human body components due to the varying motion dynamics. In this work, we adopt five independent lightweight branches to estimate the deviation of five human body components (i.e., left arm, right arm, torso, left leg, and right leg). Based on this component-wise deviation estimation strategy, we propose a Component-wise Self-Correction Network (CSCNet) to realize human motion prediction. Extensive experiments show that the CSCNet obtains state-of-the-art performance on three benchmarks, i.e., H3.6M, CMU-Mocap, and 3DPW.