CMU talk: Learning Image Manifolds != Manifold Learning

Robert Pless
Washington University in St. Louis

Monday, December 5, 2005

Abstract
This talk will detail my explorations in applying Manifold Learning techniques to real problems in image processing. Initial experiments with natural image sets (What is the intrinsic dimension of a Charlie Chaplin video clip?... Do cardio-pulmonary MR-images have a natural 2D parameterization?) illuminate several limitations of existing algorithms. First, using Euclidean (sum-of-squared pixel intensity difference) distance is usually a poor choice of image distance functions for natural images. Second, many natural image manifolds have a cyclic topology (and thus cannot be cleanly embedding into a Euclidean space). Third, natural data sets often include unlabeled examples from multiple, intersecting low-dimensional manifolds.

I will talk about several heuristic (and occasionally well founded) algorithms for choosing effective local image distance measures, finding minimal parameterizations for cyclic manifolds, and simultaneously clustering and parameterizing data from multiple intersecting manifolds. These have been brought together in an end-to-end application which automatically learns the 2D manifold structure of (ungated, free-breathing) cardiac MRI images of a patient, and uses the manifold structure of the images to regularize the segmentation of the left ventricle simultaneously in all frames.


Short Bio
Robert Pless is an Assistant Professor of Computer Science, and Assistant Director of the Center for Security Technologies at Washington University. His research interests focus on video processing; motion estimation for video surveillance and manifold learning for applications in biomedical imaging. He received a BS from Cornell University in 1994 and a PhD from the University of Maryland in 2000, and was chairman of the IEEE OMNIVIS workshop in 2003.