EDUNet: Enhancing Detail Awareness in U-Shaped Network for 2D Human Pose Estimation

Human pose estimation is crucial in computer vision for interaction, behavior analysis, and action recognition. This paper introduces EDUNet, a U-shaped key point detection network addressing information loss, redundancy, and low-resolution issues. EDUNet improves ResNet subsampling, features an enhanced BottleNeck block with partial convolution and strip pooling, removes redundancy post-upsampling, and uses a lightweight spatial grouping attention mechanism. EDUNet achieves a PCKh@0.5 of 91.6% on MPII and an AP of 76.1 on COCO, surpassing most existing models. It is robust and generalizable in real-world human body detection.