Gaussian Process Models for Human Pose and Motion

Research Overview

The inference of human shape and motion in 3D has become a topic of great interest in the vision community. The problem is difficult because people move in complex ways, having with many degrees of freedom. Their appearance is similarly hard to model due to variations in lighting, to deformations of clothing, and to occlusions. One effective way to constrain tracking is with a prior model of human pose and motion learned from motion capture data. Given the high-dimensionality of motion capture data, and the relative sparsity of data, it is natural to learn a latent probabilistic model. The Gaussian Process Latent Variable Model of Lawrence (2004) is a natural mechanism for this.

We have extended this Gaussian Process model to incorporate latent time series prediction, and to multifactor variants of the GPLVm which yield a form of style-content separation when learning multiple types of model from multiple people. We have also used this framework for monocular people tracking.

For more details see Jack Wang's GPDM page and his GPSC page.

Related Publications

Return to David Fleet's home page.