Segmentation Guided Attention Networks for Human Pose Estimation

Human pose estimation is an important and widely studied task in computer vision. One of the difficulties in human pose estimation is that the model is vulnerable to complex backgrounds when making predictions. In this paper, we propose a deep high-resolution network based on segmentation guided. A conceptually simple but computationally efficient segmentation guided module is used to generate segmentation maps. The obtained segmentation map will be used as a spatial attention map in the feature extraction stage. Since the skeletal point region is used as the foreground in the segmentation map, the model pays more attention to the key point region to effectively reduce the influence of complex background on the prediction results. The segmentation guided module provides a spatial attention map with a priori knowledge, unlike the traditional spatial attention mechanism. To verify the effectiveness of our method, we conducted a series of comparison experiments on the MPII human pose dataset and the COCO2017 keypoint detection dataset. The highest boosting effect of our model compared to HRNet on the COCO2017 dataset is up to 3%. The experimental results show that this segmentation guidance mechanism is effective in improving accuracy.