Deep High-Resolution Representation Learning for Human Pose Estimation

This paper targets human pose estimation with a focus on learning reliable high-resolution representations. Whereas most existing methods recover high-resolution representations from low-resolution ones produced by a high-to-low resolution network, the proposed network maintains high-resolution representations throughout the process: it starts from a high-resolution subnetwork, gradually adds high-to-low resolution subnetworks connected in parallel, and performs repeated multi-scale fusions so each representation repeatedly receives information from the others.

As a result the predicted keypoint heatmaps are potentially more accurate and spatially more precise, and the network demonstrates superior pose estimation on the COCO keypoint detection and MPII Human Pose datasets, plus strong pose tracking on PoseTrack. The design and its publicly released code and models became a widely used backbone for pose estimation and related dense prediction tasks.