HRNet-U: Combining the Strengths of U-Net and HRNet for Efficient Pose Estimation
Biao Guo, Fangmin Guo, Qian He · 2024
High-resolution network (HRNet) shows excellent performance in 2D multi-person pose estimation tasks. It can maintain high-resolution representation in the network and exchange information between multi-scale features, significantly improving the ability of key point prediction. However, this multiscale fusion strategy leads to computational redundancy and increases processing delays, limiting its application to resource-constrained edge devices. To solve this problem, this paper proposes a new lightweight network structure (HRNet-U), which combines the high resolution parallel advantage of HRNet and the computational efficiency of U-Net, and adopts the U-Net structure for multiscale feature fusion at each stage. Through this design, HRNet-U still maintains high resolution branch and multi-scale information, which is more suitable for the application scenarios of edge devices. This model achieves comparable or even better performance on the COCO and MPII validation sets.