SRPose: Low-Resolution Human Pose Estimation with Super-Resolution

2D human pose estimation from given images has been an activate research area in computer vision. Existing methods based on deep learning rely on high-resolution input, which is not always available in many scenarios. To address the issues, a novel algorithm called Super-Resolved Pose estimation(SRPose) is proposed in this paper, which is composed of a super-resolution sub-network(SRN) and a following human pose estimation sub-network(HPEN). The SRN equipped with global residual learning and position-preserving block constructs a HR version from a LR input and then HPEN perform pose estimation. The whole SRPose is optimized with a unified loss in end-to-ento-en. Comprehensive experiments on public benchmarks verify the effectiveness and generalization of the proposed SRPose under the condition of the LR input.