Robust scale estimation from true parameters of model

In computer vision tasks, it frequently happens that gross noise and pseudo outliers occupy the absolute majority of the data. During the past several decades, a lot of robust estimators were developed to find parameters of a model from heavily contaminated data. However, correctly estimating the parameters of a model is not enough to differentiate inliers from outliers. Robust scale estimation is often needed as the postprocessing of most robust estimators followed by a weighted least squares method on the inliers. This paper shows that the scale estimation for most robust estimators is a very weak field and more work is needed. A more robust two-step scale estimator is presented and comparative experiments show its advantages over other available scale estimatorsHanzi Wang and David Sute