Spatial and contextual aware network based on multi-resolution for human pose estimation

Aiming at capturing high-resolution spatial information and rich contextual information for accurate positioning and inference of keypoints in the task of human pose estimation, a Spatial and Contextual Aware Network (SCANet) based on multi-resolution is proposed. The network is based on HRNet and extends it with three effective modules, namely Spatial Self-Attention Module (SSAM), Information Supplement Module (ISM) and Detail Enhancement Module (DEM). The SSAM is used to provide global dependency for local features by establishing spatial correlation between locations in feature maps. ISM is proposed to further enrich spatial information and refine local representation by skip connection and dilated convolution. DEM is designed to generate high-resolution features and compensate detail information for more precise prediction. The proposed method is better than most of the state-of-the-art methods, and experiments on two keypoint benchmarks, MPII and COCO, validating the effectiveness of the model.