Occlusion-Robust 3D Hand Pose Estimation from a Single RGB Image

We propose an occlusion-robust network for 3D hand pose estimation from a single RGB image. Severe occlusions degrade the estimation accuracy of not only occluded keypoints but also visible keypoints. Since the existing methods based on a deep neural network perform convolutions on all keypoints regardless of visibility, inaccurate features from occluded keypoints affect the localization of visible keypoints. To suppress the influence of occluded keypoints, our proposed deep neural network consists of three modules: a 2D heatmap generator, parallel sub-joints network (PSJNet), and an ensemble network (EN). First, the 2D position of all keypoints in an input image is predicted as a 2D heatmap, similar to the existing methods. Then, PSJNet, which consists of several graph convolutional networks (GCN) in parallel, estimates multiple incomplete 3D poses in which some of the keypoints have been removed. Each GCN performs convolutions on a limited number of keypoints, therefore, features from occluded keypoints do not spread to the whole pose. Finally, EN merges the incomplete poses into a single 3D pose by selecting accurate positions from them. Experimental results on a public dataset RHD demonstrate that the proposed method outperforms the existing methods in the case of both small and severe occlusions.