Efficient Human Pose Estimation in Hierarchical Context

Most existing human pose estimation methods focus on enhancing the accuracy performance alone while ignoring the critical model efficiency issue. This dramatically limits their scalability and deployability in large-scale applications. In this paper, we consider the under-studied model efficiency problem in pose estimation. We demonstrate the advantages and potential of hierarchical context learning in the convolutional neural network. Specifically, we formulate a novel hierarchical context network (HCN) architecture that can be trained and deployed efficiently while achieving competitive model generalization capability. This is achieved by progressively forming and imposing multi-granularity context information during the pose regression learning process in a coarse-to-fine manner. The extensive comparative evaluations validate the superiority of the proposed HCN over a wide variety of the state-of-the-art human pose estimation models on two challenging benchmarks: MPII and LSP.