Estimate Occluded Human Pose Progressively by Graph Convolutional Network

The overlap and occlusion between people are an important factor that limits the accuracy of multi-person pose estimation. Although the previous method based on the heatmap-based approaches has achieved good results without occlusion, it is difficult to locate the coordinates of the overlap and occluded part. For this question, the direct learning of the key point coordinates based on the visible image features is ineffective and there is no reasonable use of scale information for reconstruction in the estimation process. In this article, we follow the bottom-up approach and introduce a GCN module to take advantage of this neglected information. The occlusion problem is regarded as a process of inferring implicit features from explicit features. We propose the PE-GCN module, which creatively combine the features of the key points and the connection between the key points in heat-map features and the scale features of the nodes in the pose generation process. Through self-designed evaluation metrics for invisible points in the data, our model improves the ability to locate key points of occlusion under existing benchmarks.