Image Understanding via Deformable Templates: From Representation to Inference

Michael I. Miller
Dept. of Electrical Engineering, Washington Univ.

Most real world scenes and shapes are characterized by high variability - they are not rigid - but yet they are strongly structured. While variability comes in several forms, most fundamentally it we come in all shapes and sizes. A fundamental task in the understanding and analysis of shape is therefore the construction of models that incorporate both structure as well as variability in a mathematically precise way. The global shape models introduced init Grenander's Metric Patterntheory are intended to do this. For this,it templatesare introduced, i.e. it manifoldsin 1,2, and 3 dimensions. Variability is accommodated via the introduction ofit transformationswhich act on the templates, the transformations formingit groups.

In this lecture we describe progress made on representation associated with deformable templates, as well as Bayesian inference within the parametric structures of deformable templates. We will show applications to automated target recognition and the study of the geometry of biological structures such as brains.

[Michael I. Miller, Washington Univ., Dept. of Electrical Engineering, Electronic Systems and Signals Research Laboratory, St. Louis, Missouri, 63130 USA; mim@ee.wustl.edu ]

URL http://perseus.wustl.edu/personnel/mim.html

Statistical Inference Problems in Computer Vision

Stuart Geman
Basilis Gidas
Donald McClure
Div. of Applied Mathematics, Brown Univ.

Low-level (Image Processing) and high-level (Recognition) Computer Vision problems have given rise to challenging statistical modelling and decision issues. Conceptually, both problems may be viewed as inference problems in Bayesian statistics whereby: Prior distributions represent our a priori knowledge about the world we observe, while the likelihood functions (``data models'') describe the variability of the observed grey-level data due to changes in lighting conditions, noise, blur, texturing effects, occlusion, clutter, and other degradation effects. The design of priors models is a critical and challenging step in model -based image analysis. In low-level problems, priors are typically described by Markov Random Fields (MRF); in recognition problems, priors or ``shape models'' are required to articulate contextual constraints at multiple levels. In this talk we will: (1) describe a promising approach to recognition via Hierarchical/Syntactic models (``Composition Machines'') that are closely related to Context-Free-Grammars; the approach supports dynamic programming-like computational algorithms; (2) describe some interesting issues in the parameter estimation of MRF's, due to long-range dependence and the associated phase transitions phenomena in Statistical Mechanics.