3-D hand posture recognition by training contour variation

This paper proposes a 2-D appearance-based method of estimating 3-D hand posture. The conventional methods are essentially weak in appearance changes due to the changes of 3-D postures and viewpoint. This weakness can be overcome by registering all the possible appearances but it is not suitable because of the high DOF of human hand. In the novel method, the variations of possible shape appearances (hand contour) around the registered typical appearances are trained from a number of CG images generated from 3-D hand model. The possible variations are efficiently represented as the locally-compressed feature manifold (LCFM) in an appearance feature space. The posture estimation for the sequential images is done by tracking the posture in the LCFM. Finally the experimental results show the effectiveness of the method.