The  Weizmann  Institute  of  Science
                  Faculty of Mathematics and Computer Science
 
                          Vision and Robotics Seminar
 
                  Room 141 (faculty lounge), Ziskind Building
                          on Sunday, January 15, 2012
                                    at 11:00
 
                    Note the unusual day, time and location
 
                                  David Jacobs
                             University of Maryland
 
                                 will speak on
 
                      Fast, Wavelet-Based Image Comparison
 
Abstract:
We wish to compare images in a way that allows us to consider two images to be
similar even when there are some deformations that change their shape, or
lighting variation that changes their intensity.  I will discuss our recent
work showing that we can perform image comparisons that are insensitive to
these transformations very efficiently by operating in the wavelet domain.
First, I will present an approximation algorithm for computing the Earth
Mover-Fs Distance (EMD), a metric for comparing probability distributions.
This can be used to match image descriptors, or histograms, accounting for
deformations or other variations in appearance.  Using a wavelet-based
representation, we construct an accurate, linear time algorithm for computing
the EMD. Next, we consider image matching in the presence of lighting
variation.  We attack this problem by defining a local cost on small image
changes that is insensitive to the effects of lighting variation.  We use this
cost to define a Riemannian manifold of images, in which geodesic distance is
used to measure the similarity of images when there are larger variations in
lighting.  We show that these geodesic distances can also be approximated
efficiently by computing them in the wavelet domain.   Combining these ideas,
we obtain a single image comparison method that is insensitive to lighting
changes and moderate changes in shape, and that can be computed about as
quickly as normalized correlation.  We demonstrate that this produces
state-of-the-art face recognition results in the presence of lighting and
expression changes.
 
---------------------------------------------------------
Technion Math Net-2 (TECHMATH2)
Editor: Michael Cwikel   <techm@math.technion.ac.il> 
Announcement from: Yaeli Malka   <yaeli.malka@weizmann.ac.il>