Hand Pose Estimation Using Convolutional Neural Networks and Support Vector Regression

Abstract

In order to improve the accuracy of hand pose estimation from a depth image, a method based on convolutional neural network (CNN) is proposed in this paper. First of all, we modify the structure of traditional CNN to recognize the 3D joint locations from a depth image. By appending some shortcuts between layers, the proposed network increases the correlation between the front and back layers. This structure can avoid the information loss caused by the simple layer-by-layer transmission, and can improve the estimation accuracy effectively. Afterwards, the estimated joint locations continue to be inputted into a support vector regression (SVR) phase. The use of SVR can introduce the constraint of local joint information, which can get rid of those abnormal estimations further. Extensive experiments show that our method enables significant performance improvement over the-state-of-arts in the accuracy of hand pose estimation.