Residual attention regression for 3D hand pose estimation

3D hand pose estimation is an important and challenging task for virtual reality and human-computer interaction. In this paper, we propose a simple and effective residual attention regression model for accurate 3D hand pose estimation from a depth image. The model is trained in an end-to-end fashion. Specifically, we stack different attention modules to capture different types of attention-aware features, and then implement physical constraints of the hand by projecting the pose parameters into a lower-dimensional space. In this way, 3D coordinates of hand joints are estimated directly. The experimental results demonstrate that our proposed residual attention network can achieve superior or comparable performance on three main challenging datasets, where the average 3D error is 9.7 mm on the MSRA dataset, 7.8 mm on the ICVL dataset, and 17.6 mm on the NYU dataset.