Regression Diagnostics for Multiple Model Step Data

In many vision and image problems there are multiple structures in a single data set and we need to identify the multiple models. To preserve most structures in presence of noise makes the estimation difficult. In such case for each structure, data which belong to other structures are also outliers in addition to the outliers for all the structures. Robust regression techniques are commonly used to serve the model building process for noisy data to the vision community, that fits the majority data and then to discover outliers, they tend to fail to cope with the situation. In this paper we show a newly proposed regression diagnostic measure is capable for identifying large fraction of outliers, and regression diagnostics may be a better choice to the robust regression. We demonstrate the whole thing through several artificial multiple model step data.