From volumes to views: an approach to 3-D object recognition