Pritts et al. remove user parameter from RANSAC scoring for good

Standard RANSAC variants require a user-supplied inlier scale parameter to count inliers or sum scores, which itself must be estimated from contaminated data—a fragile dependency. In their new preprint, James Pritts, Felix Seegräber, and Kevin Köser reverse the inference order: instead of estimating scale first, they marginalize it analytically in closed form under a conjugate Inverse-Gamma prior for a fixed inlier partition, then optimize over partitions. Their single expression spans non-informative Jeffreys priors and informative empirical-Bayes priors, adapting across data-rich and data-scarce regimes without altering the algorithm.

The proposed score is the first genuinely free of a scale parameter and runs in O(N log N) via sort-and-sweep. On a benchmark of nearly 70,000 image pairs covering two-view estimation with both engineered and learned features, the method exceeds state-of-the-art (RANSAC, MSAC, GaU, MAGSAC): it stays nearly flat under threshold miscalibration where baselines degrade, reaches near-optimal accuracy from just 2 validation pairs (baselines need ~100x more), and tightens prior regularization as validation data shrinks. This makes it highly practical for real-world computer vision tasks.