Dimensionality Reduction for Efficient Single Frame Hand Pose Estimation

Model based approaches for the recovery of the 3D position, orientation and full articulation of the human hand have a number of attractive properties. One bottleneck towards their practical exploitation is their computational cost. To a large extent, this is determined by the large dimensionality of the problem to be solved. In this work we exploit the fact that the parametric joints space representing hand configurations is highly redundant. Thus, we employ Principal Component Analysis (PCA) to learn a lower dimensional space that describes compactly and effectively the human hand articulation. The reduced dimensionality of the resulting space leads to a simpler optimization problem, so model-based approaches require less computational effort to solve it. Experiments demonstrate that the proposed approach achieves better accuracy in hand pose recovery compared to a state of the art baseline method using only 1/4 of the latter's computational budget.