Local Recovery of Global Eigenimages

	      H. Bischof, A. Leonardis

In this talk we propose a novel approach which combines local and
global image representations. Namely, we show how to recover the
global eigenspace representation from responses of local filter
banks. Based on that, we discuss several advantages of this {\em
global-local} approach and indicate how it can be used to considerably
enhance the current capabilities of recognition methods based on
eigenimage representations. In particular, we discuss illumination
insensitivity, selection of object specific feature points, and
invariance to plane rotations.