An Attention Module for Multi-Person Pose Estimation

In the top-down approaches of multi-person pose estimation, a human detector is adopted first to generate a set of human bounding boxes, then crop these human body and perform a single-person pose estimation model to get the final result. However, some body part of another person on the cropped image will interfere the single-person pose estimation model leading to an inaccuracy result. In order to model the relationship between adjacent keypoints effectively to alleviate this problem, we propose and attention module that could let the model get global receptive field at the shallow layer of the network and pay more attention to the key areas which is more important to pose estimation. Experiment results show that our method achieves 73.9% mAP with 2.4% absolute improvement compared to our baseline on the COCO test-dev dataset.