High-Resolution Representation Learning for Human Pose Estimation based on Transformer

Abstract

Human pose estimation requires accurate coordinate values for the prediction of human joints, which requires a high-resolution representation to effectively improve accuracy. For some difficult joint prediction tasks, it is not only necessary to look at the characteristics of the joint points themselves, but also to make judgments in combination with the context of the whole image. Generally, the resolution will be reduced when the context information is obtained. In this process, it will inevitably lose some spatial information and make the prediction inaccurate. In this paper, we propose a high-resolution human pose estimation network based on Transformer to reduce the impact of spatial information loss on keypoints estimation. In detail, we use low-level convolution neural network to extract low-level semantics from the image, and then the Transformer is used to capture the image context to further predict the key points of the human body, obtain the high-resolution representation. The experiments show that our network can accurately predict the positions of keypoints, we achieve state-of-art results on the COCO keypoint detection dataset.