UC Irvine

Accurate estimation of 3D human pose/shape from a single image remains a challenging task under occlusion or domain shift. We try to solve this problem by investigating three different geometric priors: camera pose priors, scene geometry priors, and parametric body-model priors. The first part of the dissertation focuses on analyzing the difference in the popular 3d human pose datasets and proposes a plug-in camera pose module to improve cross-dataset generalization. In the second part, we evaluate the usefulness of scene geometry in helping improve 3d pose estimators. Finally, we build strong pose/shape estimators from a single image by leveraging the best of the statistical model-based methods and nonparametric methods.