HGposeGUCN: A Lightweight Network for 3D Hand Pose Estimation from a Single RGB Image

3D hand pose estimation from a single RGB image is a challenging task which takes only one RGB image as input and predicts all the hand joints' 3D positions as output. Some previous works use two or more stages, while some other adopt some kind of model based methods to observe 3D joints positions. These previous works got relatively precise predictions, while they are always short of model weight or limited by the chosen model. In this paper, we propose a lightweight but efficient network named HGposeGUCN-Net for this problem. In our proposed HGposeGUCN-Net, we use a lightweight 2-stack hourglass network to obtain the image feature maps and 21 hand joints heatmaps first. The feature maps and initial 2D positions are fed into a Res2d module to observe the residual 2D coordination which are adopted to update the initial 2D positions got from heatmaps and bring the network more generalization ability. Finally, the refined 2D coordinates are sent into our Hand Graph U-net which can converse the 2D coordinates to 3D. The whole network can be trained end-to-end and on the published STB and RHD datasets, experiments shows that the proposed net structure has better performance.