Enhanced Human Pose Estimation with Attention-Augmented HRNet

Abstract

Human pose estimation is a pivotal task in computer vision, aiming to predict the spatial locations of key body joints within an image accurately. The challenge arises from the need to understand complex human poses, occlusions, and variations in body configurations, which often perplex traditional pose estimation models. To bolster the accuracy and robustness of human pose estimation models, we introduce an Attention-Augmented HRNet Architecture. This proposed model augments the original HRNet by integrating self-attention mechanisms. These mechanisms capture long-range dependencies among keypoints and concentrate on pivotal body regions more effectively. Experimental results demonstrate that the Attention-Augmented HRNet surpasses the baseline HRNet that lacks attention, attaining state-of-the-art performance on the COCO dataset. Specifically, our model achieves an Average Precision (AP) of 74.5%.