Pith. sign in

REVIEW 2 major objections 2 minor 24 references

RANSAC Scoring Done Right

T0 review · 2 major / 2 minor · reviewed 2026-06-29 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read RANSAC scoring can be made independent of the inlier scale by analytically marginalizing it out under a conjugate prior.

desk verdict They derive a RANSAC score by marginalizing the inlier scale analytically under an Inverse-Gamma prior, removing it from the formula, and the large benchmark shows gains in robustness. read the letter →

arxiv 2606.27385 v1 pith:XOP3YECQ submitted 2026-06-12 cs.LG cs.CV

James Pritts , Felix Seegräber , Kevin Köser This is my paper · ORCID
classification cs.LGcs.CV
keywords RANSACrobustestimationscalemarginalizationInverse-Gammamodelscoringtwo-viewgeometryinlierdetection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper shows that the usual practice of estimating an inlier scale before scoring models in RANSAC can be avoided by reversing the inference order. For any candidate set of inliers, the scale is integrated out exactly using an Inverse-Gamma prior, producing a closed-form score that depends only on the residuals and the partition. This yields a single expression that works without user-tuned thresholds and adapts to different amounts of validation data. On a large set of image pairs the resulting score outperforms earlier methods and degrades less when the threshold is set incorrectly.

What carries the argument

The central mechanism is the closed-form marginal likelihood of the inlier residuals after integrating out the scale parameter with an Inverse-Gamma prior, which is then used to score and select the best inlier partition.

What would settle it

If on the 70 000 image pair benchmark the new score shows no improvement or loses robustness when the threshold is deliberately miscalibrated, the advantage over existing scores would be refuted.

Watch

Extended reading notes

Core claim

The central discovery is a RANSAC score obtained by marginalizing the inlier scale analytically in closed form under a conjugate Inverse-Gamma prior for any fixed inlier partition and then optimizing the partition. A single formula covers both non-informative and informative priors, so the score contains no explicit dependence on inlier scale. It is computed in O(N log N) time and, on nearly 70 000 image pairs, exceeds the accuracy of RANSAC, MSAC, GaU and MAGSAC while remaining stable under threshold miscalibration and reaching near-optimal performance from only two validation pairs.

Load-bearing premise

The inlier residuals are distributed so that an Inverse-Gamma prior on the scale is conjugate and permits exact marginalization for any fixed partition.

Editorial extensions

If this is right

  • The score requires no user-supplied function of inlier scale.
  • Performance stays nearly constant even if the threshold is mis-set.
  • Near-optimal results are obtained with very small validation sets.
  • The same procedure handles both abundant and scarce validation data.
  • Efficient sorting-based computation is possible.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • This marginalization technique may apply to other robust fitting problems that involve unknown scale parameters.
  • It suggests that explicit scale estimation can be replaced by integration in many model-selection tasks.
  • In practice, the method could simplify pipelines that currently tune RANSAC thresholds separately for each dataset.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

Summary. The paper proposes a RANSAC scoring function obtained by reversing inference order: for any fixed inlier partition the inlier scale is marginalized in closed form under a conjugate Inverse-Gamma prior, producing a single parameter-free expression that spans Jeffreys and empirical-Bayes regimes. The resulting score is claimed to be the first without explicit dependence on inlier scale, to admit O(N log N) evaluation, and to outperform RANSAC, MSAC, GaU and MAGSAC on a benchmark of nearly 70 000 image pairs while remaining robust to threshold miscalibration and accurate with as few as two validation pairs.

Significance. If the conjugacy-based marginalization is exact and the reported benchmark gains hold under statistical scrutiny, the work removes a long-standing hyperparameter from robust estimation pipelines and supplies a principled way to adapt regularization strength to data availability; this would be a substantive advance for two-view geometry and related tasks.

major comments (2)
  1. [Abstract] Abstract (paragraph on reversal of inference order): the claim that the marginalization is exact and closed-form for any fixed partition requires that the inlier residuals follow a likelihood exactly conjugate to the Inverse-Gamma prior (normally Gaussian errors on variance). The manuscript must state the precise residual model and supply the derivation; without it the assertion that scale is 'genuinely absent' cannot be verified and the central contribution is at risk.
  2. [Results / Experiments] Benchmark results paragraph: superiority, flatness under miscalibration, and near-optimal accuracy from two validation pairs are asserted, yet no error bars, statistical tests, or ablation on prior hyperparameters are referenced. These omissions make it impossible to judge whether the reported gains are load-bearing for the claim of consistent outperformance.
minor comments (2)
  1. [Abstract] Typo: 'ont he order of 100 times more' should read 'on the order of 100 times more'.
  2. [Abstract] The abstract would be strengthened by stating the explicit closed-form score (or at least its functional dependence) rather than only describing its properties.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the presentation of the central contribution. We address each major point below and will incorporate the suggested changes in the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract (paragraph on reversal of inference order): the claim that the marginalization is exact and closed-form for any fixed partition requires that the inlier residuals follow a likelihood exactly conjugate to the Inverse-Gamma prior (normally Gaussian errors on variance). The manuscript must state the precise residual model and supply the derivation; without it the assertion that scale is 'genuinely absent' cannot be verified and the central contribution is at risk.

    Authors: The residual model is i.i.d. zero-mean Gaussian errors whose variance is the unknown inlier scale; the Inverse-Gamma prior is conjugate, yielding an exact closed-form marginal likelihood after integrating out the scale. The derivation appears in Section 3.2 (Eqs. 4–7) together with the explicit statement that the resulting score contains no free scale parameter. To make this immediately verifiable from the abstract, we will add a concise sentence stating the Gaussian likelihood and note that the conjugacy derivation is supplied in the main text. revision: yes

  2. Referee: [Results / Experiments] Benchmark results paragraph: superiority, flatness under miscalibration, and near-optimal accuracy from two validation pairs are asserted, yet no error bars, statistical tests, or ablation on prior hyperparameters are referenced. These omissions make it impossible to judge whether the reported gains are load-bearing for the claim of consistent outperformance.

    Authors: We agree that quantitative support for the reported gains should be strengthened. The revised manuscript will add (i) error bars computed as standard deviation across 10 independent random seeds for each method, (ii) paired Wilcoxon signed-rank tests on the 70 k pairs to establish statistical significance of the observed improvements, and (iii) an ablation table varying the Inverse-Gamma shape and rate hyperparameters over two orders of magnitude to confirm that performance remains stable. These additions directly address the concern about load-bearing evidence. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity in marginalization derivation

full rationale

The paper derives its RANSAC score by analytically marginalizing the inlier scale under a conjugate Inverse-Gamma prior for any fixed inlier partition, yielding a closed-form expression that removes scale dependence. This is a standard Bayesian technique relying on conjugacy between the assumed residual likelihood and the prior, not a self-referential definition, fitted parameter renamed as prediction, or load-bearing self-citation. The abstract explicitly states the conjugacy assumption as the enabling condition without reducing the result to its inputs by construction. No self-citations or ansatzes from prior author work are invoked in the provided text to justify the core step. The derivation is therefore self-contained against external benchmarks.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

The derivation rests on the conjugate Inverse-Gamma assumption for the scale parameter; no free parameters are introduced because the score is obtained by exact integration rather than fitting.

assumptions (1)
  • domain assumption Inlier residuals admit an Inverse-Gamma conjugate prior on the scale parameter that permits exact marginalization for any fixed partition.
    Invoked to obtain the closed-form score without estimating scale.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RANSAC Scoring Done Right." pith.science (2026). https://pith.science/paper/XOP3YECQ

@misc{pith2026260627385,
  author       = {Pith},
  title        = {Pith review of: RANSAC Scoring Done Right},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XOP3YECQ}},
  note         = {Machine review of arXiv:2606.27385}
}
read the original abstract

The most widely used RANSAC variants score candidate models by counting inliers or summing per-point scores that saturate beyond a residual threshold. Every such score requires a user-supplied parameter that is a function of the inlier scale, which must itself be estimated from contaminated data. We remove this dependence by reversing the usual order of inference: rather than estimating the scale and then scoring against it, we marginalize the inlier scale analytically in closed form under a conjugate Inverse-Gamma prior for a fixed inlier partition, then optimize over partitions. A single closed-form expression spans the non-informative Jeffreys limit and informative empirical-Bayes priors, so the same score adapts across data-rich and data-scarce regimes without any change to the algorithm. The proposed RANSAC score is the first in which the inlier scale is genuinely absent from the formula. The score admits O(N log N ) computation via sort-and-sweep. On a benchmark of nearly 70 000 image pairs spanning different two-view estimation problems and both engineered and learned feature pipelines, the proposed score exceeds the state of the art (RANSAC, MSAC, GaU, MAGSAC): it stays nearly flat under threshold miscalibration where baselines degrade, reaches near-optimal accuracy from as few as two validation pairs where baselines need ont he order of 100 times more,. and tightens its prior regularization as validation data grows scarce.

Figures

Figures reproduced from arXiv: 2606.27385 by the authors.

Figure 1
Figure 1. Profile across three datasets, one per geometric problem: PT / E (left), ETH3D / F (middle), HEB / H (right). Each panel plots, as solid lines, the median rotation error across all pairs in the dataset versus each method’s hyperparameter on a log axis, swept exhaustively over a two-decade interval. Color-matched dashed horizontal lines mark each method’s achievable performance when its hyperparameter is free to be t… view at source ↗
Figure 2
Figure 2. Sensitivity: expected test rotation error (left) and its standard deviation (right) as a function of validation-set size, for three cells spanning the geometric-problem axis—PT / E (top), ETH3D / F (middle), HEB / H (bottom). Shaded bands are 95% BCa bootstrap confidence intervals over the cross-validation trials. Our methods are nearly insensitive to validation-set size: even with as few as 2–4 validation pairs the… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 2 canonical work pages

  1. [1]

    Three things everyone should know to improve object retrieval

    Relja Arandjelovi´c and Andrew Zisserman. Three things everyone should know to improve object retrieval. InCVPR, pages 2911–2918, 2012

    2012

  2. [2]

    Progressive NAPSAC: sampling from gradually growing neighborhoods

    Daniel Barath, Maksym Ivashechkin, and Jiˇrí Matas. Progressive NAPSAC: Sampling from gradually growing neighborhoods.arXiv:1906.02295, 2019

    work page Pith review arXiv 1906

  3. [3]

    MAGSAC: Marginalizing sample consensus

    Daniel Barath, Jana Noskova, and Ji ˇrí Matas. MAGSAC: Marginalizing sample consensus. In CVPR, 2019

    2019

  4. [4]

    MAGSAC++, a fast, reliable and accurate robust estimator

    Daniel Barath, Jana Noskova, Maksym Ivashechkin, and Ji ˇrí Matas. MAGSAC++, a fast, reliable and accurate robust estimator. InCVPR, 2020

    2020

  5. [5]

    A large-scale homography benchmark

    Daniel Barath, Dmytro Mishkin, Michal Polic, Wolfgang Förstner, and Jiˇrí Matas. A large-scale homography benchmark. InCVPR, pages 21360–21370, 2023

    2023

  6. [6]

    Matching with PROSAC—progressive sample consensus

    Ondˇrej Chum and Jiˇrí Matas. Matching with PROSAC—progressive sample consensus. In CVPR, pages 220–226, 2005

    2005

  7. [7]

    Locally optimized RANSAC

    Ondˇrej Chum, Jiˇrí Matas, and Josef Kittler. Locally optimized RANSAC. InDAGM, pages 236–243, 2003

    2003

  8. [8]

    Two-view geometry estimation unaffected by a dominant plane

    Ondˇrej Chum, Tomáš Werner, and Jiˇrí Matas. Two-view geometry estimation unaffected by a dominant plane. InCVPR, pages 772–779, 2005

    2005

Show all 24 references
  1. [9]

    Starter kit and benchmark data, 2020

    CVPR 2020 RANSAC Tutorial. Starter kit and benchmark data, 2020. https://github.com/ ducha-aiki/ransac-tutorial-2020-data

    2020

  2. [10]

    SuperPoint: Self-supervised interest point detection and description

    Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-supervised interest point detection and description. InCVPRW, 2018

    2018

  3. [11]

    Fischler and Robert C

    Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography.Communications of the ACM, 24(6):381–395, 1981

    1981

  4. [12]

    RANSAC for (quasi-)degenerate data (QDEGSAC)

    Jan-Michael Frahm and Marc Pollefeys. RANSAC for (quasi-)degenerate data (QDEGSAC). InCVPR, pages 453–460, 2006

    2006

  5. [13]

    Image matching across wide baselines: From paper to practice.IJCV, 129(2), 2021

    Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiˇrí Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice.IJCV, 129(2), 2021

    2021

  6. [14]

    Fixing the locally optimized RANSAC

    Karel Lebeda, Jiˇrí Matas, and Ondˇrej Chum. Fixing the locally optimized RANSAC. InBMVC, 2012

    2012

  7. [15]

    LightGlue: Local feature matching at light speed

    Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local feature matching at light speed. InICCV, 2023

    2023

  8. [16]

    David G. Lowe. Distinctive image features from scale-invariant keypoints.IJCV, 60(2):91–110, 2004

    2004

  9. [17]

    D. R. Myatt, P. H. S. Torr, S. J. Nasuto, J. M. Bishop, and R. Craddock. NAPSAC: High noise, high dimensional robust estimation. InBMVC, pages 458–467, 2002

    2002

  10. [18]

    Schönberger, Viktor Larsson, and Marc Pollefeys

    Johannes L. Schönberger, Viktor Larsson, and Marc Pollefeys. Fixing the RANSAC stopping criterion.arXiv:2503.07829, 2025

    2025

  11. [19]

    Schönberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger

    Thomas Schöps, Johannes L. Schönberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. InCVPR, 2017

    2017

  12. [20]

    RANSAC scoring functions: Analysis and reality check.Preprint, submitted to IJCV, 2025

    Alexander Shekhovtsov. RANSAC scoring functions: Analysis and reality check.Preprint, submitted to IJCV, 2025. 11

    2025

  13. [21]

    Robust computation and parametrization of multiple view relations

    Philip Torr and Andrew Zisserman. Robust computation and parametrization of multiple view relations. InICCV, pages 727–732, 1998

    1998

  14. [22]

    Torr and Andrew Zisserman

    Philip H. Torr and Andrew Zisserman. MLESAC: A new robust estimator with application to estimating image geometry.Comput. Vis. Image Underst., 78(1):138–156, 2000

    2000

  15. [23]

    Philip H. S. Torr. Bayesian model estimation and selection for epipolar geometry and generic manifold fitting.IJCV, 50(1):35–61, 2002. 12 Supplementary Material A Model Dimensions: Parameters, Codimension, and Minimal Sample Size Table 1 summarizes the key dimensional quantiti...

    2002

  16. [24]

    inlier noise

    Crucially, GaU requires σ (and hence τ) to be supplied externally, whereas our method jointly estimates ˆsand selects the optimal partition from the data (Section 3.2). Solving (16) forτ ∗ in closed form: τ ∗ = ˆs q 2 log a p 2/π /ˆs .(18) For typical values (e.g., a= 10ˆs), t...

Pith tools

Reviewed June 29, 2026 · model on record in the stance chip above.