Robust Regression: a Computer Vision Perspective
To exhibit a robust behavior, data analysis techniques incorporate a priori information into the procedure. Most often this information is related to the largest tolerable scale of the inlier noise, or equivalently, to the smallest percentage of inliers. The success of the analysis is contingent upon the correctness of the assumptions. We show that all the robust regression techniques popular in computer vision are special cases of the general class of M-estimators with auxiliary scale. Next, we reinterpret M-estimation within the projection pursuit framework, and describe a new robust regression technique whose performance is much less conditioned on the accuracy of additional information. Affine motion and fundamental matrix estimation, from data in which the percentage of outliers significantly exceeds that of inliers, are used as examples. Time permitting, we will also discuss the issue of data containing multiple structures.
Related papers are available at:
PETER MEER received the Dipl. Engn. degree from the Bucharest Polytechnic Institute, Bucharest, Romania, and the D.Sc. degree from the Technion, Israel Institute of Technology, Haifa, Israel, both in electrical engineering. In 1991 he joined the Department of Electrical and Computer Engineering, Rutgers University and is currently a Professor. His research interest is in application of modern statistical methods to image understanding problems.