Estimating 3D Human Pose from Silhouette Image using Approximate Chamfer Distance and Kernel Subspace

摘要

We propose a novel method for estimating 3D human pose from silhouette image. The proposed method includes two major parts: obtaining candidate poses and then re-ranking them. For a query image, a certain number of candidate poses are retrieved efficiently from a two million poses database using a novel approximate chamfer distance. Then, Kernel CCA based re-ranking algorithm assists to improve the rank order of candidate poses. We have demonstrated the effectiveness of our method on synthetic images involving large pose variations.