Scale-Aware Network with Attentional Selection for Human Pose Estimation

Human pose estimation is a fundamental yet challenging task in computer vision. Human pose estimation from a single image is a challenging problem due to the limited information of 2D images and the large variations in configuration and appearance of body parts. Recent works has largely improved the result of human pose estimation because of the development of convolutional neural network. However, there still exists many difficult cases, such as occluded keypoints, complex background and scale variations of human body keypoints, which cannot be well dealt with. In this paper, we design a novel scale-aware network with attentional selection that extracts multi-scale semantic information and meaningful features. Specifically, we propose a Feature Pyramid Supervision Module (FPSM), which can improve the estimation accuracy of scale variations. Meanwhile, a Spatial and Channel Attention Module (SCAM) is designed for recalibrating the spatial and channel features. Based on the proposed algorithm, we achieve state-of-the-art result on LSP dataset and make competitive performance on MPII Human Pose dataset.