Human Pose Estimation Based on Light-Weight High-Resolution Network with Polarized Self-Attention

Abstract

With the development of computer vision technology, human pose estimation as an indispensable part of human-computer interaction has gradually attracted the attention of researchers. Unlike most existing high-to-low resolution networks in human pose estimation networks, HRNet can maintain high-resolution representations throughout the process, resulting in reliable output results. However, HRNet has the problem of large amount of operation parameters and loss of feature information when multi-resolution feature fusion occurs. Therefore, based on HRNet-32, this paper makes improvements and proposes Human Pose Estimation Based on Light-Weight High-Resolution Network with Polarized Self- Attention. By introducing the improved Gaff module of GhostNet into the first stage of HRNet-32, the network is successfully lightened, and the spatial and channel feature information is extracted more carefully through the Polarized Self-Attention mechanism, and a more accurate human pose estimation effect is obtained.