Manifold learning and stylization of human motion

Abstract

In this thesis, we present two motion generative models for constructing motion manifold and motion stylization. First, we propose a novel recurrent neural network-based method to construct a latent motion manifold that can represent a wide range of human motions in a long sequence. Our model create compact and versatile motion manifold that allows for generating new motions by performing random sampling and algebraic operations, such as interpolation and analogy, in the latent motion manifold. Second, we propose the motion style transfer model that can control the motion style of individual body parts. Inputs to our framework are one source motion for content, and multiple target motions, each for stylizing a different body part, and its output is a content-preserving motion stylized by body part. The advantages of our framework are thus it can generate an extensive variety of motions by stylizing them by body part and can allows style interpolation by body part, which further enhances its usefulness. In addition, our framework can be easily integrated with other motion generator. We demonstrate a real-time motion style transfer by integrating with an existing motion controller.