Fast Human Pose Retrieval Using Approximate Chamfer Distance

The estimation of 3D human pose from a single image can be implemented in the way of large-scale image retrieval. For a given input image, a few similar images are retrieved from the database consisting of human figure images annotated with 3D human poses; then the 3D poses corresponding to retrieved images serve as the pose estimates. This method is simple but works if two conditions are met: (i) sufficient data and (ii) a good image matching algorithm. Sufficient data can be generated by using 3D character rendering software and various human motion data. As for matching algorithm, here we employ the chamfer distance which has proved to be an effective tool in many related works. However, applying the chamfer distance to large-scale problem would lead to high time requirements. Thus, here we propose an efficient approximate chamfer distance which uses the subspace representation of Distance Transform (DT) in computing chamfer distance so that the major computation can be done offline. We show in experiments that the approximate chamfer distance achieved competitive estimation performance but over three hundred times speedup gain in comparison with the exact chamfer distance.