3D Human Pose Reconstruction Using Millions of Exemplars

We propose a novel exemplar based method to estimate 3D human poses from single images by using only the joint correspondences. Due to the inherent depth ambiguity, estimating 3D poses from a monocular view is a challenging problem. We solve the problem by searching through millions of exemplars for optimal poses. Compared with traditional parametric schemes, our method is able to handle very large pose database, relieves parameter tweaking, is easier to train and is more effective for complex pose 3D reconstruction. The proposed method estimates upper body poses and lower body poses sequentially, which implicitly squares the size of the exemplar database and enables us to reconstruct unconstrained poses efficiently. Our implementation based on the kd-tree achieves real-time performance. The experiments on a variety of images show that the proposed method is efficient and effective.