Aiming at the challenge of modeling the relations globally among different body joints. In this paper, we proposed a simple network based on graph reasoning for human pose estimation, which named SP-GRe. We introduce dilated convolution to construct a Dilated Bottleneck Module (DRM), which can enlarge the receptive field and exploit its feature extraction capability. Meanwhile, it can enhance the model’s local representations of each key point. In view of the potential advantages of graph-based propagation, we design a Global Graph Reasoning module (GGR) based on graph convolution. The module stores the explicit joints in the graph structure for global relationship reasoning. By aggregating the features of local joints and global graph nodes, GGR enables the accurate location of key points in the interaction between projection and back-projection. Comprehensive experiments demonstrate that the proposed method achieves superior top-down pose estimation results on two benchmark datasets, MSCOCO and MPII. Moreover, SP-GRe demonstrates superior results on human pose estimation over popular human pose estimation networks.