An Improved Concat Module Based on HRNet for 2D Multi-person Pose Estimation

The existing shallow networks cannot solve the problem of insufficient cross between feature information and insufficient semantic information for 2D human pose estimation task, while deepening network will lead to gradient disappearance or over-fitting. In this paper, we propose an Improved Concat Module (ICM) based on the HRNet network to improve the accuracy and robustness of 2D multi-person pose estimation. The proposed ICM module is composed of tandem skip connection units and average pooling operations in parallel, where the input information of each improved skip connection unit performs element-wise product operation with the convolution branch, which improves the crossover between feature information more effectively and can reduce the gradient disappearance or over-fitting. Furthermore, three improved skip connection units are cascaded in this paper to provide rich semantic information for the final regression of human pose key points and enhance the nonlinear expression ability of the network. In this paper, our method achieves 70.1%AP score on the COCO test-dev dataset and 69.2%AP score on the CrowPose test dataset, which is significantly improved over multiple state-of-the-art methods and achieves robust results in the case of multiple people.