Multi-path spatial attention for human pose estimation in complex scenes

In human pose estimation in complex scenes, the complexity of background and foreground increases, and the existing attention mechanism is generally unable to effectively capture the spatial position information of human keypoints. In this paper, an effective multi-path spatial attention (MSA) is proposed, which consists of group normalization convolution (GNCov) and multi-path spatial learning (MSL). GNConv can process different feature groups independently, making the model more flexible in dealing with complex backgrounds and foregrounds. The core of MSL consists of three sets of depthwise separable convolutions in parallel, which can effectively extract spatial position relationships. By comparing with the state-of-the-art attention, MSA has a better ability to focus on key areas in complex scenes. Our code is publicly available at https://github.com/ruanhaoyuan/MSA/tree/master.