Modeling Variation in Motion Data

We present a new method to model and synthesize variation in human motion. Given a small amount of input motion data, we learn a generative model that is able to synthesize new output motion variations that are statistically similar to the input data. The new variations retain the features of the original data examples, but are not exact copies. Our model does not require timewarping or synchronization of similar sequences of motions. We learn a Dynamic Bayesian Network model from the input data that enables us to capture properties of conditional independence in the data, and build a multivariate probability distribution of it. We present synthesis results across a range of different types of motions, and demonstrate novelty and aesthetic appeal of the new variations generated with respect to the input motions. Our technique can synthesize new motions efficiently and has a small memory requirement.