Human pose estimation using patch-based candidate generation and model-based verification

We present an algorithm for 3d pose estimation of articulated people in natural images. The poses are disassembled into a collection of local patches and a new pose is inferred by assembling the local patches. This concept allows inference of a wide variety of poses from a small number of training patches. The actual process is realized efficiently by a novel voting scheme where each local patch extracted from the input image is matched to the model patches and matchings cast votes for possible locations and poses of the human body, yielding a set of candidate location-pose pairs. Each candidate is then holistically verified using a top-down model based method, where SVM regression computes the final score by aggregating several scores capturing different features of the candidates. We evaluate our method on both real and synthetic images and demonstrate its ability.