Hite-HRNet:A High-Performance High-Resolution Human Pose Estimation Network

In human pose estimation, high-resolution networks have demonstrated remarkable multi-scale feature extraction capabilities, achieving impressive results. However, balancing computational efficiency and performance remains challenging, especially in resource-constrained environments. We propose a novel lightweight high-resolution network, Hite-HRNet. This network combines the advantages of global context modeling and multi-scale feature fusion. Specifically, we introduce an enhanced channel-spatial attention method, implemented through two functional modules: Global Pose Attention (GPA) and Multi-Scale Convolution for Pose Estimation (MSC). These modules are embedded into the stem and IterativeHead, respectively, to improve the network’s ability to represent human pose and small targets, such as keypoints, while maintaining low resource consumption. Our work highlights the potential of multi-scale features in advancing pose estimation performance. Experimental results demonstrate that Hite-HRNet achieves performance on the COCO dataset that is competitive with state-of-the-art methods, enhancing pose estimation accuracy while maintaining low resource demands.