* Faculty       * Staff       * Contact       * Institute Directory
* Research Groups      
* Undergraduate       * Graduate       * Institute Admissions: Undergraduate | Graduate      
* Events       * Institute Events      
* Lab Manual       * Institute Computing      
No Menu Selected

* Research

Ph.D. Theses

Robust Region Extraction: Extracting model and domain parameters in the presence of noise and multiple populations

By Amitha Perera
Advisor: Charles V. Stewart
May 5, 2003

A fundamental component of many computer vision problems is extracting a region of interest from a data set and estimating the parameters of a model describing the data in the region. It occurs in diverse situations such as 3-D reconstruction from range data; computer-aided cartography; tracking; object recognition; and feature-based registration. In an important class of problems, there is some data model and region shape information available for the region of interest, but no other information is available for the rest of the data. We present an approach for solving this class of problems. Dubbed the DBM-Estimator, it is a framework that combines all available model and shape constraints to jointly estimate the parameters of a data model and region boundary that describes the region of interest.

Our formulation combines M-estimation and region growing, and poses the region extraction problem as the joint minimization of two inter-dependent cost functions. Starting from a small initial region, the DBM-Estimator evolves the region to the final estimate, while simultaneously estimating the data model parameters during the region evolution. The distinguishing feature of the DBM-Estimator is that it only requires a data model for the region of interest, and is thus ``non-competitive''. The DBM-Estimator is controlled primarily by only two parameters, which simplifies the application to various problems.

The data model and region parameter estimation is robust, and thus the DBM-Estimator can cope with outliers and modeling errors. Theoretical analysis shows that the DBM-Estimator is relatively insensitive to the parameter settings, which allows the DBM-Estimator to be applied to a variety of data sets without hand-tuning. Experimental comparisons with mixture models and the mean shift on synthetic data shows that the DBM-Estimator is able to detect extremely subtle discontinuities that the other approaches cannot.

We demonstrate the DBM-Estimator on three diverse problems: surface extraction from range data, forest boundary detection from satellite images, and vasculature extraction from 3-D CTA data.

* Return to main PhD Theses page