Motionlet LLC coding for discriminative human pose estimation

Abstract

3D human pose estimation is a challenging but important research topic with abundant applications. As for discriminative human pose estimation, the main goal is to learn a nonlinear mapping from image descriptors to 3D human pose configurations, which is difficult due to the high-dimensionality of human pose space and the multimodality of the distribution. To address these problems, we propose a novel motionlet LLC coding in a discriminative framework. A motionlet consists of training examples covering a local area in terms of image space, pose space and time stream. We first group most informative and helpful training examples into motionlets, then perform LLC Coding to learn the nonlinear mapping and get candidate poses, and finally choose the most appropriate pose as the result estimate. To further eliminate ambiguities and improve robustness, we extend our framework to incorporate multiviews. We conduct qualitative evaluation on our Taichi data set and quantitative evaluation on HumanEva data set, which show that our approach has gained the-state-of-the-art performance and significant improvement against previous approaches.