Abstract
•ICR is an approach for geometry-based clustering that requires no threshold, even in the presence of multiple clusters.•r-ICR is a recursive formulation of ICR which does not require the number of clusters to be externally specified.•Validation of both ICR variants on real datasets for motion segmentation.•Unlike baselines, ICR and r-ICR time complexity grows more slowly as a function of number of points.
We present an approach for motion clustering based on a novel observation that a signature for putative pixel correspondences can be generated by collecting their residuals with respect to model hypotheses drawn randomly from the data. Inliers of the same motion cluster should have strongly correlated residuals, which are low when a hypothesis is consistent with the data in the cluster and high otherwise. After evaluating a number of hypotheses, members of the same cluster can be identified based on these correlations. Due to this property, we named our approach Inlier Clustering based on the Residuals of Random Hypotheses (ICR). An important advantage of ICR is that it does not require an inlier-outlier threshold or parameter tuning. In addition, we propose a supervised recursive formulation of ICR (r-ICR) that, unlike many motion clustering methods, does not require the number of clusters to be known a priori, as long as annotated data are available for training. We validate ICR and r-ICR on several publicly available datasets for robust geometric model fitting.