An Efficient Method for Boosting Human Pose Estimation

Abstract

Most existing human pose estimation approaches fall into designing new network architectures or tend to apply deeper layers. Most of the methods are time-consuming, and lightweight plug-in for improving human pose estimation to the existing network gets little investigation. In this paper, we propose a lightweight plug-in module to boost the performance of human pose estimation named PoseReNed. PoseReNet is a network with three branches that are designed to tackle the attenuation caused by occluded keypoints or different scales of the keypoints. Generally, small-scale keypoints are more difficult to detect, and we observe that different channels of the output feature map have different attributes to the performance of estimation. We apply a channel attention mechanism to re-weight the channel to trade-off among different scales of the keypoints. By aggregating multiscale output feature maps, the pose estimation performance can be improved. Serving as a model-agnostic plug-in, PoseReNet brings about significant performance boost to existing human pose estimation models. Extensive experiments show that PoseReNet can effectively improve precision on COCO and MPII.