MERCREDI 1er OCTOBRE, AMPHI RATAUD, 16H

David Jacobs (djacobs@umiacs.umd.edu)
University of Maryland
http://www.cs.umd.edu/~djacobs/



Title: Matching Images with Deformations


Abstract: When we match images that come from the same object, we must
often allow for 2D, non-linear deformations.  These can model changes
in shape that can occur when an object deforms or an articulated
object moves its parts, differences in shape between different
instances of the same type of object, or variations in apparent shape
due to changes in viewpoint.  This talk will provide an overview of
several approaches to matching that stress finding problem
formulations that yield computationally efficient algorithms.  First,
we present an approximation algorithm for computing the Earth Mover’s
Distance, a metric for comparing probability distributions that can be
used to match image descriptors, accounting for deformations.  Using a
wavelet-based representation, we construct an accurate, linear time
algorithm. Next, we describe a novel image descriptor that is
invariant to deformations, and a shape descriptor that is invariant to
articulations. We describe the use of this shape matching algorithm in
a hand-held device for computer assisted plant species identification.
Finally, we show that the cost returned by a stereo matching algorithm
can be used for object recognition, without using 3D models or
performing reconstruction.  We use this to construct a face
recognition algorithm that compares 2D gallery and probe images taken
from different viewpoints.  This algorithm outperforms all prior work
on the CMU PIE data set for face recognition with pose variation.