Combination of Positions and Angles for Hand Pose Estimation.

This paper deals with the estimation of hand pose from a single depth image. We present a method that is based on a description of the hand pose via local rotations of bones trained discriminatively in an end-to-end fashion using a convolutional neural network. We compare our method with existing approach of hand pose estimation of 3D locations of hand joints. For this purpose, we collected precise ground-truth data with a passive marker-based optical motion capture technology. The results show, that the estimation of the hand pose formulated as a combination of local rotations of bones and relative locations of joints outperforms the direct estimation of 3D global joints locations.