A-HRNet: Attention Based High Resolution Network for Human pose estimation
Ying Li, Chenxi Wang, Yu Cao, Benyuan Liu, Yan Luo, Honggang Zhang
Abstract
Recently, human pose estimation has received much attention in the research community due to its broad range of application scenarios. Most architectures for human pose estimation use multiple resolution networks, such as Hourglass, CPN, HRNet, etc. High Resolution Network (HRNet) is the latest SOTA architecture improved from Hourglass. In this paper, we propose a novel attention block that leverages a special Channel-Attention branch. We use this attention block as the building block and adopt the architecture of HRNet to build our Attention Based HRNet (A-HRNet). Experiments show that our model can consistently outperform HRNet on different datasets. Moreover, our model achieves the state-of-the-art performance on the COCO keypoint detection val2017 dataset (77.7 AP) <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> .