Posterior constraints for double-counting problem in clustered pose estimation

In this paper, we propose a novel and integrated framework to estimate human pose. Firstly, a pose cluster of the relative location between connected parts is applied before pictorial structure modeling, which can make each model more faithful and the whole estimation more flexible to various kinds of poses. And then, different from previous single global model, we propose the mixture pictorial structure models based on the clusters to obtain the parts candidates. Furthermore, to overcome the double-counting problem, we also present a constraint function to recombine the candidates derived from the optimal clustered model. Experiments on a publicly challenging dataset show that our method is more accurate and flexible and performs effectively in tackling the double-counting phenomena.