Learning spectral transform for 3D human motion prediction

Abstract

In existing motion prediction methods that use graph convolution networks, motion sequences are transformed to a spectral domain, and future motions are predicted through graph spectral filtering for the transformed spectral sequences. However, because the conventional spectral transform method uses a predetermined spectral basis, the motion prediction does not work well for aperiodic or complicated motions. To overcome this problem, we propose a method to learn spectral domain transforms from motion sequences in the training dataset. To this end, two methods are attempted: one for learning the frequency of each spectral basis, and another for learning the values of the basis function directly. Through experiments on representative 3D human motion benchmarks, H3.6M and CMU Mocap, we demonstrate that both of the proposed methods consistently outperform the baseline method. In particular, the method of directly learning the basis function outperforms the state-of-the-art methods. We also demonstrate that the proposed method yields realistic predictions, even for aperiodic and complicated action categories.