FT-RANSAC: Towards robust multi-modal homography estimation

As the golden standard in robust estimation, theclassic RANSAC approach has undergone extensive research thatcontributed to further enhancements in run-time performance,robustness, and multi-structure support to name a few. Yet, theaccelerating growth of multi-modal co-registered datasetsrequires a new adaptation of the RANSAC algorithm. In thispaper, we propose a multi-modal fault-tolerant extension toRANSAC, termed FT-RANSAC, with a model-independenttolerance to degenerate configurations. Besides building on stateof-the-art RANSAC variants, such as PROSAC, our approachintroduces a Hough inspired dimensionality reduction andconsistency voting processes, to enable robust estimation in thepresence of non-homogenous multi-modal correspondence sets.Through experimental evaluation using homography estimation ofRGB-D data, we demonstrate that our approach outperforms theclassic single-modality RANSAC in robustness and tolerance todegenerate configurations. Finally, the proposed approach lendsitself to parallel multi-core implementations, and could be adaptedto specialized RANSAC extensions found in the literature.