On the sample consensus robust estimation paradigm: comprehensive survey and novel algorithms with applications.
Olukanmi, Peter Olubunmi.
MetadataShow full item record
This study begins with a comprehensive survey of existing variants of the Random Sample Consensus (RANSAC) algorithm. Then, five new ones are contributed. RANSAC, arguably the most popular robust estimation algorithm in computer vision, has limitations in accuracy, efficiency and repeatability. Research into techniques for overcoming these drawbacks, has been active for about two decades. In the last one-and-half decade, nearly every single year had at least one variant published: more than ten, in the last two years. However, many existing variants compromise two attractive properties of the original RANSAC: simplicity and generality. Some introduce new operations, resulting in loss of simplicity, while many of those that do not introduce new operations, require problem-specific priors. In this way, they trade off generality and introduce some complexity, as well as dependence on other steps of the workflow of applications. Noting that these observations may explain the persisting trend, of finding only the older, simpler variants in ‘mainstream’ computer vision software libraries, this work adopts an approach that preserves the two mentioned properties. Modification of the original algorithm, is restricted to only search strategy replacement, since many drawbacks of RANSAC are consequences of the search strategy it adopts. A second constraint, serving the purpose of preserving generality, is that this ‘ideal’ strategy, must require no problem-specific priors. Such a strategy is developed, and reported in this dissertation. Another limitation, yet to be overcome in literature, but is successfully addressed in this study, is the inherent variability, in RANSAC. A few theoretical discoveries are presented, providing insights on the generic robust estimation problem. Notably, a theorem proposed as an original contribution of this research, reveals insights, that are foundational to newly proposed algorithms. Experiments on both generic and computer-vision-specific data, show that all proposed algorithms, are generally more accurate and more consistent, than RANSAC. Moreover, they are simpler in the sense that, they do not require some of the input parameters of RANSAC. Interestingly, although non-exhaustive in search like the typical RANSAC-like algorithms, three of these new algorithms, exhibit absolute non-randomness, a property that is not claimed by any existing variant. One of the proposed algorithms, is fully automatic, eliminating all requirements of user-supplied input parameters. Two of the proposed algorithms, are implemented as contributed alternatives to the homography estimation function, provided in MATLAB’s computer vision toolbox, after being shown to improve on the performance of M-estimator Sample Consensus (MSAC). MSAC has been the choice in all releases of the toolbox, including the latest 2015b. While this research is motivated by computer vision applications, the proposed algorithms, being generic, can be applied to any model-fitting problem from other scientific fields.