Hand Pose Estimation with Attention-and-Sequence Network.

Hand pose estimation from depth images is an essential topic in computer vision. Despite the recent advancements in this area promoted by Convolutional Neural Network, accurate hand pose estimation is still a challenging problem. In this paper, we analyse the spatial relationship among hand joints, and discover that: (1) there exists independence of joints from different fingers, and (2) there also exists strong correlation among adjacent joints in the same finger. Based on this, we present a novel Attention-and-Sequence Network (ASNet) embedded with finger attention and joint sequence mechanisms. Here the finger attention mechanism is proposed to ensure the independence of joints from different fingers, while the joint sequence mechanism is employed to make use of strong correlation among adjacent joints in the same finger. The proposed ASNet achieves an average 3D error of 5.6mm on ICVL, 10.3mm on NYU, 7.3mm onMSRA, and these competitive results further confirm the great effectiveness of ASNet.