Robust model fitting in pattern recognition

In this paper, we introduce a robust framework for model based parameter estimation. The framework is developed for a particular class of problems: one that contains many interesting examples in the computer vision area. The scheme leads to a new M-estimator, which, is shown to be optimal, in some sense, for this class of problems. This contrasts with previous works: these have employed least squares (known to be undermined by outliers), or have used an M-estimator without analysis to determine the robustness or efficiency of the estimator for the problem setting at hand. We illustrate the utility of this estimator and compare it with several previously employed M-estimators in the context of conic fitting