Learning Feature Transformations to Recognize Faces Rotated in Depth

We present a method for recognizing objects (faces) on the basis of just one stored view, in spite of rotation in depth. The method is not based on the construction of a three-dimensional model for the object. Our recognition results represent a signi cant improvement over a previous system developed in our laboratory. We achieve this with the help of a simple assumption about the transformation of local feature vectors with rotation in depth. The parameters of this transformation are learned on training examples.