Scale estimation in multiple models fitting via consensus clustering

Abstract

This paper presents a new procedure for fitting multiple geometric structures without having a priori knowledge of scale. Our method leverages on Consensus Clustering, a single-term model selection strategy relying on the principle of stability, thereby avoiding the explicit tradeoff between data fidelity (i.e., modeling error) and model complexity. In particular we tailored this model selection to the estimate of the inlier threshold of T-linkage, a fitting algorithm based on random sampling and preference analysis. A potential clustering is evaluated based on a consensus measure. The crucial inlier scale ϵ is estimated using an interval search. Experiments on synthetic and real data show that this method succeeds in finding the correct scale.