Spatio-temporal graph-based generative adversarial networks for character motion style transfer

Abstract

This paper presents a novel deep learning-based framework for the automatic generation of stylistic variations on character motion. Our framework is a single set of generative adversarial networks that learns stylistic features from a collection of unpaired motion clips with style labels while supporting multiple cross-domain mapping. We construct a spatio-temporal graph to model a motion sequence and employ graph convolution networks (GCN) to extract stylistic properties along spatial and temporal dimensions. Through spatio-temporal modeling, our framework can perform a robust style transfer on long sequential heterogeneous data and between (extremely different) independent actions. For the motion style translation task, we first use a network that maps a random noise to style, which allows diverse stylization results to be generated without using reference motion. Through various experiments, we demonstrate the ability of our method to generate improved results in terms of visual quality, stylistic diversity, and content preservation.