Multiple Model Fitting by Evolutionary Dynamics

We propose a novel multiple model fitting method based on outlier insensitive evolutionary dynamics, fulfilling several important requirements. Our method automatically identifies a unspecified number of models and is robust to noise and outliers in the data. Furthermore, we are able to handle overlapping models, by allowing that data points are assigned to more than one model. This is implicitly handled during model fitting and not as a post-processing step. Gross outliers are directly identified, by letting some points unassigned. We also introduce a technique, considering nearest neighbor analysis, to significantly reduce computation time, while maintaining model fitting accuracy. We show experiments on real-world and synthetic data, achieving accurate model fitting results also demonstrating an application of plane fitting on a consumer hardware providing RGB-D video streams.