Reconsideration of multi-stage deep network for human pose estimation

Abstract

ABSTRACT Estimating human pose is a difficult task due to the high flexibility of joints and possible occlusion. In this paper, the proposed architecture efficiently predicts human pose and solves the preceding problem. The proposed framework has three main consecutive parts: (1) Deep Convolutional Neural Network (DCNN) based feature extraction, (2) feature refinement, and (3) the fusion of detection and context information. During the feature extraction phase, we have proposed a fusion of two DCNN modules, which have been inspired by VGG-19 and Inception-v4 deep learning architectures. In the feature refinement, a cascaded feature integration technique has been proposed over the stacked hourglass, to make the system efficiently locate the challenging joints. At last, a fusion of context information with the detected prediction is performed, which makes the system accurate towards occlusion. In this way, poses with difficult joint coordinates can be reliably estimated even in the presence of occlusion or severe distracting factors. The successful testing of the proposed method has been done on popular MPII and LSP datasets. Based on the experimental results and the analysis of the selected datasets, it is found that the proposed framework is more accurate compared to other state-of-the-art methods in terms of the PCK metric.