User-Independent Robust Statistics for Computer Vision

Ph.D. Thesis Sushil Mittal


Abstract

The goal of robust methods in computer vision is to extract all the information necessary to solve a given task while discarding everything that is not needed. The tasks can be very simple or very complex, but in real-life applications, a robust procedure is always required. In the end, the performance of a machine for solving a vision problem will be judged against that of human observers performing the equivalent task. Since we know that the human visual system works in a much more sophisticated manner than the present day computer vision systems, this ultimate goal is still far away. Nonetheless, the aim of robust computer vision systems has always been to emulate human vision-like behavior in the presence of noise. Furthermore, the robust algorithms should be independent of user inputs up to quite a large extent. However, contrary to this, almost all state-of-the-art robust estimation algorithms are dependent on the user for providing some information about the underlying characteristics of the data on which the algorithm operates. Often times, it is hard for the user to supply such information to the algorithm.

The work presented here focuses on developing robust algorithms for computer vision, that can estimate the underlying model in the data without any sort of user intervention. We present several interesting applications both in geometric computer vision and medical imaging. In the first part, we present a completely user-free robust regression algorithm called the generalized projection based M-estimator (gpbM) which can estimate multiple inlier structures present in the data also containing a lot of gross outliers without any user input. We also show how the model estimate can be further refined by using optimization on Grassmann manifolds. In the second part, we present three important applications in medical imaging involving 3D computed tomography (CT) data -- automatic detection of coronary lesions, automatic correction of coronary centerlines and automatic segmentation of coronary vessels. Finally, in the third part, we present the application of automatic and robust document image alignment and comparison.

The thesis contains 144 pages.


Return to Theses