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 is estimated using an interval search. Experiments on synthetic and real data show that this method succeeds in finding the correct scale.

Multiple structure recovery with T-linkage

“…In this paper we wrap up and review all the material that appeared in [7,8,9], including more thorough descriptions, additional insights and new experiments. In particular: the background material and the description of T-Linkage have been enhanced; a new discussion on the scale estimation problem is presented; new experiments have been added, including tests on the complete Adelaide dataset and a 3D plane fitting experiment.…”