Multiple structure recovery with T-linkage

This work addresses the problem of robust fitting of geometric structures to noisy data corrupted by outliers. An extension of J-linkage (called T-linkage) is presented and elaborated. T-linkage improves the preference analysis implemented by J-linkage in term of performances and robustness, considering both the representation and the segmentation steps. A strategy to reject outliers and to estimate the inlier threshold is proposed, resulting in a versatile tool, suitable for multi-model fitting “in the wild”. Experiments demonstrate that our methods perform better than J-linkage on simulated data, and compare favorably with state-of-the-art methods on public domain real datasets.