Human Pose Estimation Method Based on Flexible Model and Deep Learning

This paper(1) proposes a mixture model that combines joint appearance-based and inter-joint spatial relationship-based models with a deep neural architecture called deep convolutional neural network (DCNN). This method has been applied to tackle the human pose estimation problem. Firstly we construct a graphical model for the human body. Secondly, the images are decomposed into several image patches which are used as positive input samples. And finally, a DCNN network that can solve multiple classifications is obtained to perform human pose estimation. The quantitative results obtained from the experiments are impressive and show that our method outperforms recent works of the state-of-the-art that used same considered datasets.