EDite-HRNet: Enhanced Dynamic Lightweight High-Resolution Network for Human Pose Estimation

Abstract

Lightweight pose estimation models have been widely used in devices with different computing powers, providing convenience for numerous downstream tasks, such as gait estimation, behavior analysis, motion capture, etc. Although these lightweight methods can run on low-performance equipment, their estimation accuracy is low, which seriously affects the actual experience. In order to improve the prediction accuracy of the lightweight human pose estimation methods, we propose an Enhanced Dynamic Lightweight High-Resolution Network (EDite-HRNet) for human pose estimation. Specifically, we propose an Enhanced Dynamic Multi-scale Context (EDMC) block which enhances the features of the simple branch with multi-level features of the complex branch to realize multi-level features fusion. Moreover, inspired by GhostNet V2, we redesign the Enhanced Dynamic Global Context (EDGC) and the Enhanced Dynamic Multi-scale Context (EDMC) block by adopting GhostNet V2 module with DFC attention to replace ConvBN block in the original blocks. The experimental results on the two datasets (66.1% on the COCO2017 dataset and 86.8% on the MPII dataset), demonstrate that our network achieves the state-of-the-art performance with a slight increase in model complexity.