Human Pose Prediction by Progressive Generation in Multi-scale Frequency Domain

We address a problem of 3D human pose prediction from a sequence of human body skeletons. To model the spatio-temporal dynamics, the discrete cosine transform (DCT) and the graph convolutional networks (GCN) are often applied to signals on a human skeleton graph. By DCT, temporal information of a human skeleton sequence can be embedded into the frequency domain. However, in previous studies, the prediction models using DCT implicitly learned each frequency coefficient by gradients calculated from a loss of the predictions and the ground truths of human body skeletons. In this paper, we propose a progressive human pose prediction model in frequency domain so that explicitly predict high-, medium-, and low-frequency motion of a target person. We confirmed that the proposed method improves prediction accuracy through experiments using public datasets on Human3.6M and CMU Mocap datasets.