A gradient-based eigenspace approach to dealing with occlusions and non-Gaussian noise

In the recent literature, gradient-based (filtered) eigenspaces have been used as a means to achieve illumination insensitivity. In this paper we show that filtered eigenspaces are also inherently robust w.r.t. (non-Gaussian) noise and occlusions. We argue that this robustness stems essentially from the sparseness of representation and insensitivity w.r.t. shifts in the mean value. This is also demonstrated experimentally using examples from the field of object recognition and pose estimation.