Synthesis and Editing of Human Motion with Generative Human Motion Model

In this paper, a generic approach is presented to constructing generative motion model from prerecorded motion data that allows the synthesis of new motions or modification of existing motions in various ways. The key idea is to decompose human motion data into a series of latent variables which decompose a number of sports from different properties, in the way motion variations are interpreted by modeling human motion data in the Gaussian process. The effectiveness and flexibility of this approach in experiments and applications are demonstrated by constructing two generative motion models as an example of various kinds of motion properties.