Human Pose Estimation using DirectionMaps

In this paper, we propose a novel approach to detect human pose in a wild image. The approach uses a nonparametric representation, which we refer as DirectionMaps, to learn the direction information of human body parts. The whole architecture is divided into two stages. The first stage is designed to jointly learn parts location and direction of each body part relative to the center of body. The second stage integrates the location and direction information to obtain more accurate posture. Our Convolutional Neural Network(CNN) can learn both location and direction information, which is better able to detect human pose. We evaluated our method on the MPH Human Pose datasets, reaching a good performance.