Year: 2025
Type: conference-paper
Abstract: Deep learning has been widely used in the field of human pose estimation, and has demonstrated remarkable performance on a wide range of datasets. High resolution images have been widely used to improve the performance of human pose estimation. However, it is also necessary to deal with model cost and computation complexity. To solve this problem, an efficient and lightweight high resolution network named Elite-HRNet is proposed. The network achieves performance improvement through an innovative lightweight architecture and optimized feature extraction capabilities. It replaces traditional high-resolution network structures with the Contextual Capture Stem Block and Conditional Channel Spatial Weighting Block, while proposing a novel multi-branch parallel architecture based on high resolution to achieve cross-resolution multi-scale data fusion. Experimental results demonstrate that the lightweight Elite-HRNet achieves outstanding performance with only 0.17% of parameters, attaining AR of 68.8 and AP of 74.6 on the COCO dataset, along with an average accuracy of 86.7% on the MPII dataset.
Authors: Hao Ma, Bin Shao, Jingnan Dong
Institution: Huzhou Normal University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4410956726