Partial Attention CenterNet for Bottom-Up Human Pose Estimation

Abstract

The typical bottom-up human pose estimation methods can be divided into two steps, keypoint detection and grouping. The traditional keypoint regression-based methods exploit an effective backbone (like HRNet) and different prediction heads to acquire the body center and body joint. Then they utilize the offset between the body center and body joint to figure out the grouping task. In this paper, we first propose a body branch module and keypoint attention module to improve keypoint detection and keypoint regression. In body branch module, we exploit a multi-branch structure for keypoint detection and keypoint regression. Each branch represents a part of human body. In keypoint attention module, two simple yet reliable pooling layers are adopted to extract the attention areas of different kinds of keypoints. Combining these two modules, we propose a Partial Attention CenterNet for multi-person human pose estimation. The proposed method outperforms the traditional keypoint regression-based methods. Experiments have demonstrated the obvious performance improvements on COCO dataset brought by the introduced components.