Hand pose estimation with multi-scale network

Hand pose estimation plays an important role in human-computer interaction. Because it is a problem of high-dimensional nonlinear regression, the accuracy achieved by the existing methods of hand pose estimation are still unsatisfactory. With the development of deep neural networks, more and more people have begun to adopt the method involving deep neural network.We proposed a multi-scale convolutional neural network for the single depth image of the hand. The network, which is end-to-end, directly calculates the three-dimensional coordinates of the joints of the hand,and the multi-scale structure enhances the convergence speed and the output accuracy of the network. In addition, an output function for the output layer, called Stair Rectified Linear Units, is used to limit the output value. As a result of experiments, the optimization method with momentum is found not suitable for hand pose estimation because it is a task of unstable regression. Finally our proposed method has state-of-the-art performance on the NYU Hand Pose Dataset.