Fixing the RANSAC Stopping Criterion

Abstract: For several decades, RANSAC has been one of the most commonly used robust estimation algorithms for many problems in computer vision and related fields. The main contribution of this paper lies in addressing a long-standing error baked into virtually any system building upon the RANSAC algorithm. Since its inception in 1981 by Fischler and Bolles, many variants of RANSAC have been proposed on top of the same original idea relying on the fact that random sampling has a high likelihood of generating a good hypothesis from minimal subsets of measurements. An approximation to the sampling probability was originally derived by the paper in 1981 in support of adaptively stopping RANSAC and is, as such, used in the vast majority of today's RANSAC variants and implementations. The impact of this approximation has since not been questioned or thoroughly studied by any of the later works. As we theoretically derive and practically demonstrate in this paper, the approximation leads to severe undersampling and thus failure to find good models. The discrepancy is especially pronounced in challenging scenarios with few inliers and high model complexity. An implementation of computing the exact probability is surprisingly simple yet highly effective and has potentially drastic impact across a large range of computer vision systems.

Analysis of "Fixing the RANSAC Stopping Criterion"

The paper "Fixing the RANSAC Stopping Criterion" by Johannes Schönberger, Viktor Larsson, and Marc Pollefeys, addresses a critical oversight in the widely utilized Random Sample Consensus (RANSAC) algorithm. RANSAC, originally introduced by Fischler and Bolles in 1981, has been instrumental in solving various robust estimation tasks within computer vision, such as motion segmentation, stereo vision, and structure-from-motion. Despite RANSAC's popularity, a fundamental flaw in its stopping criterion persists throughout its numerous adaptations, impacting the algorithm's reliability and efficiency.

Key Contributions

The central thesis of the paper is the identification and rectification of an imprecise approximation used to determine the probability of selecting an all-inlier sample within RANSAC's iterations. The original RANSAC algorithm approximates this probability, leading to potential undersampling, especially in scenarios with low inlier ratios or high model complexity. The authors theoretically derive the exact probability, revealing the extent of the approximation error, and propose a robust solution that corrects this longstanding issue.

Numerical Findings and Practical Implications

The study presents an in-depth numerical analysis, illustrating that the traditional approximation method results in significant undersampling errors. These errors are particularly pronounced in challenging scenarios where the inlier ratio is low, and the model's degree of freedom is considerable. By implementing the exact probability calculation, the paper demonstrates a marked improvement in model estimation across synthetic and real-world datasets. For instance, in synthetic 2D geometric primitive fitting experiments, the correct application of the stopping criterion consistently yields better models, reflected by substantial enhancements in the area-under-the-curve (AUC) metrics at various accuracy thresholds.

Impact Across Computer Vision Systems

Correcting the RANSAC stopping criterion may significantly enhance the performance of a vast array of computer vision systems. The fix is notably straightforward to implement, demanding minimal computational overhead, yet offers substantial improvements in robust model estimation accuracy. The revised stopping condition impacts a plethora of RANSAC-based algorithms like LO-RANSAC, PROSAC, and MAGSAC, among others, and extends to applications in structure-from-motion, camera localization, and multi-model fitting.

Future Directions

By addressing a fundamental error in RANSAC's adaptive stopping method, this paper potentially resets the standard for robust estimation in computer vision. Future research could focus on integrating this fix within various RANSAC-based frameworks, examining its effects in more complex and diverse scenarios. Furthermore, combining this stopping criterion revision with recent advances in adaptive sampling and inlier re-weighting strategies might result in even more powerful estimation techniques tailored to specific computer vision applications.

In summary, this paper significantly contributes to the field of computer vision by refining a foundational algorithm employed in numerous applications. The findings encourage a reevaluation of RANSAC implementations, urging the community to adopt this enhancement for more accurate and reliable model estimation.