RANSAC Scoring Done Right
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "RANSAC Scoring Done Right".
Jane: Here's the gist of the paper. The most widely used RANSAC variants score candidate models by counting inliers or summing per-point scores that saturate beyond a residual threshold <ref:2606.27385#pg2> Every such score requires a user-supplied parameter that is a function of the inlier scale, which must itself be estimated from contaminated data <ref:2606.27385#pg2> The proposed RANSAC score is the first in which the inlier scale is genuinely absent from the formula <ref:2606.27385#pg2> This score admits O(N log N) computation via sort-and-sweep <ref:2606.27385#pg2> 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 <ref:2606.27385#pg2> The Core Innovation The key idea is to reverse the usual elimination order <ref:2606.27385#pg2> The scores above first remove the latent inlier/outlier labels (by summation in the marginal likelihood or by maximization in the profile likelihood), leaving a score that still depends on a quantile of σ <ref:2606.27385#pg2> We instead marginalize σ analytically under a conjugate Inverse-Gamma prior given a fixed inlier partition; the non-informative Jeffreys prior arises as a limiting case of this Inverse-Gamma family, so the same closed form covers both informative and non-informative priors <ref:2606.27385#pg2> A single score therefore adapts to data availability without any change to the algorithm: an empirical-Bayes prior fit from a small validation set when labelled data exists, and the Jeffreys limit, which needs no validation data at all, when it does not <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> Crucially, a is not a relabelled inlier scale: it parameterizes the outlier component of the mixture and is not tied by any quantile relation to the inlier scale <ref:2606.27385#pg2> After marginalization the inlier scale is genuinely absent from the formula; this is the structural sense in which our score is the first in the RANSAC family to properly marginalize σ, and, where the per-pair inlier count is large enough for the data to dominate the prior, the score is markedly less sensitive to a than threshold-based scores are to their hyperparameter <ref:2606.27385#pg2> Derivation of Scale Invariance For a fixed partition I we marginalize σ under a conjugate Inverse-Gamma prior given a fixed inlier partition; the non-informative Jeffreys prior arises as a limiting case of this Inverse-Gamma family, so the same closed form covers both informative and non-informative priors <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> Marginalizing σ squared under the conjugate Inverse-Gamma prior admits a closed form; accounting for the nθ parameters consumed by fitting θ (which reduce the effective degrees of freedom from nI dg to ν = nI dg − nθ) and taking logarithms yields the scale-free marginal score (Section 7) <ref:2606.27385#pg2> The best model maximizes S jointly over θ and I <ref:2606.27385#pg2> Model Adaptation and Ranking The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> Empirical Validation We validate these claims at scale on nearly 70 000 image pairs spanning homography, essential-, and fundamental-matrix estimation with both engineered (RootSIFT) and learned (SuperPoint+LightGlue) feature pipelines (Section 4) <ref:2606.27385#pg2> The score matches the best-tuned threshold-based baselines at their optima while being far more robust to threshold miscalibration where the data informs the scale and reaching near-optimal accuracy from as few as two validation pairs where baselines need ∼100× more <ref:2606.27385#pg2> Sensitivity Analysis The proposed methods reach near-optimal expected test error at validation budgets as small as V = 2, while the threshold-based baselines need ∼100× more validation data to converge <ref:2606.27385#pg2> On PhotoTourism the three IG variants (Jeffreys, IG-MLE, IG-Robust) sit within 0.006◦ of their large-V asymptote of 0.77◦ already at V = 2, whereas MSAC, GaU, and MAGSAC++ start at ∼0.90◦ and need V ≈ 1024 to reach the same floor <ref:2606.27385#pg2> For the practitioner, validation-size insensitivity means a handful of ground-truth pairs (or none, in the data-rich regime where Jeffreys suffices) is enough to deploy near-optimally, which distinguishes the proposed family from the threshold-based state of the art <ref:2606.27385#pg2> To our knowledge, this is the first RANSAC score that genuinely marginalizes the noise scale σ in closed form <ref:2606.27385#pg2> References
1: Relja Arandjelovic and Andrew Zisserman.
Tom: Three things everyone should know to improve object retrieval.
Jane: In CVPR, pages 2911–2918, 2012 <ref:2606.27385#pg6>
2: Daniel Barath, Maksym Ivashechkin, and Jiˇrí Matas.
Tom: Progressive NAPSAC: Sampling from gradually growing neighborhoods.
Paper summary: Tom: So to wrap up "RANSAC Scoring Done Right," we're talking about a scoring framework that genuinely eliminates the inlier scale dependence in RANSAC scoring, which is something they say is structurally different from previous methods.
Jane: It’s about taking the usual approach where you estimate the noise scale and then score against it, and instead they reverse that order by marginalizing sigma analytically under an Inverse-Gamma prior.
Lu: They claim this leads to a single closed-form expression that covers both informative empirical Bayes priors and the non-informative Jeffreys limit, which means it adapts across data regimes without needing to change the underlying algorithm.
Meng: The implication for us is that if we have limited ground truth pairs, we don't need huge validation sets just to tune a parameter related to the noise scale; this score lets us deploy near optimally with very little data.
Lalam: They also show that this score admits an O of n log n computation via sort and sweep, and they’ve shown it performs better than threshold based baselines under miscalibration.
Jane: The authors say this is the first RANSAC score in which the inlier scale is genuinely absent from the formula, which they find very significant for how we think about model selection in practice.
Tom: It moves us away from needing a user supplied parameter that has to be estimated from contaminated data, which simplifies our workflow significantly.
Conclusion: Tom: So we've been looking at this paper, "RANSAC Scoring Done Right," and what’s really striking is how they tackle the core problem of estimating that noise scale, which used to be a huge headache in RANSAC.
Jane: It sounds like they found a way to get around needing that extra parameter that you have to guess from noisy data.
Lu: Exactly. The main technical move here is taking the usual scoring method and instead marginalizing the noise scale analytically using an inverse gamma prior, which makes the score scale-free.
Meng: Scale-free means it doesn't care about some arbitrary number we pick for how much noise there is; it figures out that information itself. But does that mean we don't need any validation data at all?
Jane: That’s a big part of the claim, actually. They suggest that if you have enough data, or even just a few pairs, you can get near-optimal results without needing hundreds of validation samples like before.
Tom: That’s what they show in their validation experiments. They compare it against all these other methods—MSAC, GaU, MAGSAC++—and this new score stays flat even when the threshold settings are slightly off, which is a huge win for deployment reliability.
Lu: It’s about structural invariance. The paper claims this is the first RANSAC score where you genuinely get rid of that noise scale dependence in closed form, and they show it can run in big O of n log n time, which is pretty efficient for large datasets.
Jane: So for someone listening who isn't deep into math, the implication is that you can use this score with much less hassle when you’re trying to find those good matches.
Meng: From my side at the startup, if it means I can ship a model that doesn't break because of how I tuned one hyperparameter, then yeah, that’s worth looking at. But we need to see if that O of n log n is actually fast enough for our real traffic.
Tom: Right. It’s about making the whole pipeline more robust so it just works better in the real world without us having to constantly fine-tune every little setting. And next up, we're going to look at how they actually ran these tests on different kinds of image matching problems.
James Pritts, Felix Seegräber, Kevin Köser
Marine Data Science Institute for Informatics, Kiel University
cs.LG, cs.CV
Submitted: 2026-06-12
Updated: 2026-11-10
Comments: Camera-ready version, NeurIPS 2026. 20 pages
Journal ref: NeurIPS 2026
Code: https://github.com/ducha-aiki/ransac-tutorial-2020-data
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 76/100
The gist: The most widely used RANSAC variants score candidate models by counting inliers or summing per-point scores that saturate beyond a residual threshold <ref:2606.27385#pg2> Every such score requires a…
Key concepts
- Inlier Scale ($\sigma$)
- This is a parameter that represents the typical scale of the inliers in a dataset. In traditional RANSAC scores, this scale must be estimated from the contaminated data, which introduces dependency and potential bias. The paper's innovation is to remove this explicit dependence from the final score.
- Marginalization under Inverse-Gamma Prior
- Instead of estimating $\sigma$ directly, the method marginalizes it analytically using a conjugate Inverse-Gamma prior. This mathematical step integrates out the unknown scale parameter, resulting in a closed-form 'scale-free' score that is independent of $\sigma$. This allows the score to work reliably even when validation data is scarce.
- Scale-Free Marginal Score
- This is the final proposed scoring function. It is a closed-form expression involving residual sum of squares, inlier count, and the outlier domain half-width ($a$). Crucially, it does not depend on the estimated inlier scale ($\sigma$), meaning it adapts naturally to data quality and availability without needing external hyperparameter tuning.
Terminology
Summary
The most widely used RANSAC variants score candidate models by counting inliers or summing per-point scores that saturate beyond a residual threshold <ref:2606. p. 0 de kk@informatik. de
Abstract
The most widely used RANSAC variants score candidate models by counting
inliers or summing per-point scores that saturate beyond a residual threshold. de
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. A single closed-form expression spans the non-informative Jeffreys
limit and informative empirical-Bayes priors, so the same score adapts across datarich 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 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. 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 ∼100× more,
and tightens its prior regularization as validation data grows scarce.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The Core Innovation
The key idea is to reverse the usual elimination order <ref:2606.27385#pg2> The scores above first remove the latent inlier/outlier labels (by summation in the marginal likelihood or by maximization in the profile likelihood), leaving a score that still depends on a quantile of σ <ref:2606. p. 1 The key idea is to
reverse the usual elimination order. The scores above first remove the latent inlier/outlier labels
(by summation in the marginal likelihood or by maximization in the profile likelihood), leaving a
score that still depends on a quantile of σ. The scores above first remove the latent inlier/outlier labels
(by summation in the marginal likelihood or by maximization in the profile likelihood), leaving a
score that still depends on a quantile of σ. We instead marginalize σ analytically under a conjugate
Inverse-Gamma prior given a fixed inlier partition; the non-informative Jeffreys prior arises as a
limiting case of this Inverse-Gamma family, so the same closed form covers both informative and
non-informative priors. We instead marginalize σ analytically under a conjugate
Inverse-Gamma prior given a fixed inlier partition; the non-informative Jeffreys prior arises as a
limiting case of this Inverse-Gamma family, so the same closed form covers both informative and
non-informative priors. A single score therefore adapts to data availability without any change to
the algorithm: an empirical-Bayes prior fit from a small validation set when labelled data exists, and
the Jeffreys limit, which needs no validation data at all, when it does not. We then optimize over
partitions, which requires only a sort over residuals. The resulting scale-free marginal score is a closed-form expression involving the residual sum of
squares, the inlier count, and the outlier domain half-width a (Section 3), and admits O(N log N)
computation via a sort-and-sweep algorithm. We then optimize over
partitions, which requires only a sort over residuals. The resulting scale-free marginal score is a closed-form expression involving the residual sum of
squares, the inlier count, and the outlier domain half-width a (Section 3), and admits O(N log N)
computation via a sort-and-sweep algorithm. We then optimize over
partitions, which requires only a sort over residuals. The resulting scale-free marginal score is a closed-form expression involving the residual sum of
squares, the inlier count, and the outlier domain half-width a (Section 3), and admits O(N log N)
computation via a sort-and-sweep algorithm. We then optimize over
partitions, which requires only a sort over residuals. The resulting scale-free marginal score is a closed-form expression involving the residual sum of
squares, the inlier count, and the outlier domain half-width a (Section 3), and admits O(N log N)
computation via a sort-and-sweep algorithm. We then optimize over
partitions, which requires only a sort over residuals. The resulting scale-free marginal score is a closed-form expression involving the residual sum of
squares, the inlier count, and the outlier domain half-width a (Section 3), and admits O(N log N)
computation via a sort-and-sweep algorithm. We then optimize over
partitions, which requires only a sort over residuals. The resulting scale-free marginal score is a closed-form expression involving the residual sum of
squares, the inlier count, and the outlier domain half-width a (Section 3), and admits O(N log N)
computation via a sort-and-sweep algorithm. We then optimize over
partitions, which requires only a sort over residuals. The resulting scale-free marginal score is a closed-form expression involving the residual sum of
squares, the inlier count, and the outlier domain half-width a (Section 3), and admits O(N log N)
computation via a sort-and-sweep algorithm. The resulting scale-free marginal score is a closed-form expression involving the residual sum of
squares, the inlier count, and the outlier domain half-width a (Section 3), and admits O(N log N)
computation via a sort-and-sweep algorithm. Crucially, a is not a relabelled inlier scale: it parameterizes the outlier component of the mixture and is not tied by any quantile relation to the inlier
scale. Crucially, a is not a relabelled inlier scale: it parameterizes the outlier component of the mixture and is not tied by any quantile relation to the inlier
scale. After marginalization the inlier scale is genuinely absent from the formula; this is the structural
sense in which our score is the first in the RANSAC family to properly marginalize σ, and, where
the per-pair inlier count is large enough for the data to dominate the prior, the score is markedly less
sensitive to a than threshold-based scores are to their hyperparameter.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Derivation of Scale Invariance
For a fixed partition I we marginalize σ under a conjugate Inverse-Gamma prior given a fixed inlier partition; the non-informative Jeffreys prior arises as a limiting case of this Inverse-Gamma family, so the same closed form covers both informative and non-informative priors <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> Marginalizing σ squared under the conjugate Inverse-Gamma prior admits a closed form; accounting for the nθ parameters consumed by fitting θ (which reduce the effective degrees of freedom from nI dg to ν = nI dg − nθ) and taking logarithms yields the scale-free marginal score (Section 7) <ref:2606. p. 3 2 The Scale-Free Marginal Score
Marginalizing σ
2 under the conjugate Inverse-Gamma prior admits a closed form; accounting for the
nθ parameters consumed by fitting θ (which reduce the effective degrees of freedom from nI dg to
ν = nI dg − nθ) and taking logarithms yields the scale-free marginal score (Section C gives the full
derivation):
S(θ, I) = log Γν
2
+ α0
−
ν
2
+ α0
logRSSI
2
+ β0
−
ν
2
log(2π) − nO dg log(2a), (7)
where ν = nI dg − nθ and RSSI =
P
i∈I qi with qi = ∥ri∥
2
the per-point squared residual. The
best model maximizes S jointly over θ and I. 2 The Scale-Free Marginal Score
Marginalizing σ
2 under the conjugate Inverse-Gamma prior admits a closed form; accounting for the
nθ parameters consumed by fitting θ (which reduce the effective degrees of freedom from nI dg to
ν = nI dg − nθ) and taking logarithms yields the scale-free marginal score (Section C gives the full
derivation):
S(θ, I) = log Γν
2
+ α0
−
ν
2
+ α0
logRSSI
2
+ β0
−
ν
2
log(2π) − nO dg log(2a), (7)
where ν = nI dg − nθ and RSSI =
P
i∈I qi with qi = ∥ri∥
2
the per-point squared residual. The
best model maximizes S jointly over θ and I.Where the paper says it
Where the paper says it
Model Adaptation and Ranking
The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2> The resulting scale-free marginal score is a closed-form expression involving the residual sum of squares, the inlier count, and the outlier domain half-width a (Section 3) and admits O(N log N) computation via a sort-and-sweep algorithm <ref:2606.27385#pg2>
Empirical Validation
We validate these claims at scale on nearly 70 000 image pairs spanning homography, essential-, and fundamental-matrix estimation with both engineered (RootSIFT) and learned (SuperPoint+LightGlue) feature pipelines (Section 4) <ref:2606.27385#pg2> The score matches the best-tuned threshold-based baselines at their optima while being far more robust to threshold miscalibration where the data informs the scale and reaching near-optimal accuracy from as few as two validation pairs where baselines need ∼100× more <ref:2606. p. 1 We validate these claims at scale on nearly 70 000 image pairs spanning homography, essential-, and
fundamental-matrix estimation with both engineered (RootSIFT) and learned (SuperPoint+LightGlue)
feature pipelines (Section 4). The score matches the best-tuned threshold-based baselines at their
optima while being far more robust to threshold miscalibration where the data informs the scale and
reaching near-optimal accuracy from as few as two validation pairs, where the baselines need ∼100×
more.Where the paper says it
Sensitivity Analysis
The proposed methods reach near-optimal expected test error at validation budgets as small as V = 2, while the threshold-based baselines need ∼100× more validation data to converge <ref:2606.27385#pg2> On PhotoTourism the three IG variants (Jeffreys, IG-MLE, IG-Robust) sit within 0. p. 8 1 Results
The proposed methods reach near-optimal expected test error at validation budgets as small as
|V| = 2, while the threshold-based baselines need ∼100× more validation data to converge. On
PhotoTourism the three IG variants (Jeffreys, IG-MLE, IG-Robust) sit within 0. On
PhotoTourism the three IG variants (Jeffreys, IG-MLE, IG-Robust) sit within 0. 006◦ of their large-|V|
asymptote of 0. 006◦ of their large-|V|
asymptote of 0. 77◦
already at |V| = 2, whereas MSAC, GaU, and MAGSAC++ start at ∼0. 77◦
already at |V| = 2, whereas MSAC, GaU, and MAGSAC++ start at ∼0. 90◦
and need |V| ≈ 1024 to reach the same floor; vanilla RANSAC starts higher (1.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
For the practitioner, validation-size insensitivity means a handful of ground-truth pairs (or none, in the data-rich regime where Jeffreys suffices) is enough to deploy near-optimally, which distinguishes the proposed family from the threshold-based state of the art <ref:2606.27385#pg2> To our knowledge, this is the first RANSAC score that genuinely marginalizes the noise scale σ in closed form <ref:2606. p. 9 For the practitioner,
validation-size insensitivity means a handful of ground-truth pairs (or none, in the data-rich regime
where Jeffreys suffices) is enough to deploy near-optimally, which distinguishes the proposed family
from the threshold-based state of the art. To our knowledge, this is the first RANSAC score that
genuinely marginalizes the noise scale σ in closed form.Where the paper says it
References
[1] Relja Arandjelovic and Andrew Zisserman. Three things everyone should know to improve object retrieval. p. 0 MAGSAC [3] and MAGSAC++ [4] were introduced specifically to remove this dependence by
“marginalizing” over σ under a uniform prior σ ∼ U[0, σ¯]. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[2] Daniel Barath, Maksym Ivashechkin, and Jiˇrí Matas. Progressive NAPSAC: Sampling from gradually growing neighborhoods. p. 4 Crucially τ
∗ adapts per hypothesis through sˆ,
whereas the baselines apply one global τ ; in the high-ν limit our score is the Bayes-optimal Gaussian–
Uniform ranking objective and threshold-based scoring its σ-fixed specialization (Section E). 4 Experiments
State-of-the-art RANSAC pipelines achieve performance through the combined effect of multiple
components: guided sampling (PROSAC [6], NAPSAC [17]), degeneracy testing (DEGENSAC [8],
QDEGSAC [12]), local optimization (LO-RANSAC [7, 14]), and varied termination criteria [18]. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
Where the paper says it
[3] Daniel Barath, Jana Noskova, and Jiˇrí Matas. MAGSAC: Marginalizing sample consensus. p. 0 We call scores of this form threshold-based. MAGSAC [3] and MAGSAC++ [4] were introduced specifically to remove this dependence by
“marginalizing” over σ under a uniform prior σ ∼ U[0, σ¯]. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[4] Daniel Barath, Jana Noskova, Maksym Ivashechkin, and Jiˇrí Matas. MAGSAC++, a fast, reliable and accurate robust estimator. p. 5 Shekhovtsov [20] showed that the four threshold-based scoring functions we compare—
RANSAC [11], MSAC [21], GaU [22], and MAGSAC++ [4]—can each be reparameterized with
a single hyperparameter. Our three proposed marginal-likelihood variants (Jeffreys, IG-Robust,
IG-MLE) likewise share a single hyperparameter, the outlier half-width from (7). Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[5] Daniel Barath, Dmytro Mishkin, Michal Polic, Wolfgang Förstner, and Jiˇrí Matas. A large-scale homography benchmark. p. 1 After marginalization the inlier scale is genuinely absent from the formula; this is the structural
sense in which our score is the first in the RANSAC family to properly marginalize σ, and, where
the per-pair inlier count is large enough for the data to dominate the prior, the score is markedly less
sensitive to a than threshold-based scores are to their hyperparameter. We validate these claims at scale on nearly 70 000 image pairs spanning homography, essential-, and
fundamental-matrix estimation with both engineered (RootSIFT) and learned (SuperPoint+LightGlue)
feature pipelines (Section 4). Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[6] Ondˇrej Chum and Jiˇrí Matas. Matching with PROSAC—progressive sample consensus. p. 4 Crucially τ
∗ adapts per hypothesis through sˆ,
whereas the baselines apply one global τ ; in the high-ν limit our score is the Bayes-optimal Gaussian–
Uniform ranking objective and threshold-based scoring its σ-fixed specialization (Section E). 4 Experiments
State-of-the-art RANSAC pipelines achieve performance through the combined effect of multiple
components: guided sampling (PROSAC [6], NAPSAC [17]), degeneracy testing (DEGENSAC [8],
QDEGSAC [12]), local optimization (LO-RANSAC [7, 14]), and varied termination criteria [18]. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[7] Ondˇrej Chum, Jiˇrí Matas, and Josef Kittler. Locally optimized RANSAC. p. 0 RANSAC Scoring Done Right
James Pritts Felix Seegräber Kevin Köser
Marine Data Science, Institut für Informatik
Kiel University, Germany
jpr@informatik. de fse@informatik. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[8] Ondˇrej Chum, Tomáš Werner, and Jiˇrí Matas. Two-view geometry estimation unaffected by a dominant plane. p. 0 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 ∼100× more,
and tightens its prior regularization as validation data grows scarce. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[9] CVPR 2020 RANSAC Tutorial. Starter kit and benchmark data, 2020 https://github. p. 0 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 ∼100× more,
and tightens its prior regularization as validation data grows scarce. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-supervised interest point detection and description. p. 4 This matches the GaU threshold at an effective prior γeff =
1
2
, so the noninformative γ =
1
2
(Section 3. 1) is self-consistent. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[11] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. p. 8 1: 1 000 minimal-sample models are drawn per pair, scored at every hyperparameter κ, the
best-scoring model is selected at each κ, and its error against the GT pose is recorded. With these
errors cached, the bootstrap loop is fast: draw a validation pool V of cardinality |V| (on HEB, sampled
from a random subset of scenes held out for validation, with the disjoint remaining scenes forming the
test pool, so trial variance reflects cross-scene generalization; split at the pair level on PhotoTourism
and ETH3D) and pick the validation-optimal hyperparameter κ
⋆
(V) = arg minκ medianp∈V ep(κ),
and evaluate e
test(κ
⋆
(V)) = medianp∈T ep(κ
⋆
(V)) on the held-out test pool T . Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[12] Jan-Michael Frahm and Marc Pollefeys. RANSAC for (quasi-)degenerate data (QDEGSAC). p. 4 Crucially τ
∗ adapts per hypothesis through sˆ,
whereas the baselines apply one global τ ; in the high-ν limit our score is the Bayes-optimal Gaussian–
Uniform ranking objective and threshold-based scoring its σ-fixed specialization (Section E). 4 Experiments
State-of-the-art RANSAC pipelines achieve performance through the combined effect of multiple
components: guided sampling (PROSAC [6], NAPSAC [17]), degeneracy testing (DEGENSAC [8],
QDEGSAC [12]), local optimization (LO-RANSAC [7, 14]), and varied termination criteria [18]. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[13] 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 <ref:2606. p. 0 Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
[14] Karel Lebeda, Jiˇrí Matas, and Ondˇrej Chum. Fixing the locally optimized RANSAC. p. 0 RANSAC Scoring Done Right
James Pritts Felix Seegräber Kevin Köser
Marine Data Science, Institut für Informatik
Kiel University, Germany
jpr@informatik. de fse@informatik. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[15] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local feature matching at light speed. p. 1 After marginalization the inlier scale is genuinely absent from the formula; this is the structural
sense in which our score is the first in the RANSAC family to properly marginalize σ, and, where
the per-pair inlier count is large enough for the data to dominate the prior, the score is markedly less
sensitive to a than threshold-based scores are to their hyperparameter. We validate these claims at scale on nearly 70 000 image pairs spanning homography, essential-, and
fundamental-matrix estimation with both engineered (RootSIFT) and learned (SuperPoint+LightGlue)
feature pipelines (Section 4). Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[16] David G. Lowe. Distinctive image features from scale-invariant keypoints. p. 1 After marginalization the inlier scale is genuinely absent from the formula; this is the structural
sense in which our score is the first in the RANSAC family to properly marginalize σ, and, where
the per-pair inlier count is large enough for the data to dominate the prior, the score is markedly less
sensitive to a than threshold-based scores are to their hyperparameter. We validate these claims at scale on nearly 70 000 image pairs spanning homography, essential-, and
fundamental-matrix estimation with both engineered (RootSIFT) and learned (SuperPoint+LightGlue)
feature pipelines (Section 4). Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
[17] D. R. Myatt, P. H. S. Torr, S. J. Nasuto, J. M. Bishop, and R. Craddock. NAPSAC: High noise, high dimensional robust estimation <ref:2606. p. 4 Crucially τ
∗ adapts per hypothesis through sˆ,
whereas the baselines apply one global τ ; in the high-ν limit our score is the Bayes-optimal Gaussian–
Uniform ranking objective and threshold-based scoring its σ-fixed specialization (Section E). 4 Experiments
State-of-the-art RANSAC pipelines achieve performance through the combined effect of multiple
components: guided sampling (PROSAC [6], NAPSAC [17]), degeneracy testing (DEGENSAC [8],
QDEGSAC [12]), local optimization (LO-RANSAC [7, 14]), and varied termination criteria [18].Where the paper says it
[18] Johannes L. Schönberger, Viktor Larsson, and Marc Pollefeys. Fixing the RANSAC stopping criterion. p. 0 RANSAC Scoring Done Right
James Pritts Felix Seegräber Kevin Köser
Marine Data Science, Institut für Informatik
Kiel University, Germany
jpr@informatik. de fse@informatik. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
Where the paper says it
[19] 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. p. 0 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 ∼100× more,
and tightens its prior regularization as validation data grows scarce. Shekhovtsov [20] deconstructs the
derivation and identifies three errors (an inlier density that assigns zero likelihood to the ground-truth
model, an event probability equated with a density value, and IRLS weights chosen first and integrated
1Code will be released upon acceptance. arXiv:2606.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Marginalizing noise scale analytically removes dependence on inlier scale, allowing a single score to adapt across data-rich and data-scarce regimes
without any change to the algorithm.
This enables a system to perform reliably with as few as two validation pairs where baselines need∼100× more,
achieving near-optimal accuracy. -
The sort-and-sweep algorithm provides O(N log N) computation, enabling fast scoring of candidate models by evaluating the score over sorted per-point scores, which is a significant computational improvement over methods that require exhaustive search or complex label marginalization sums.
-
The proposed score's implicit threshold adapts to the noise level of each hypothesis;
A well-fitting model with small inlier residuals produces a tight sˆ and hence a tight τ∗; a poor model with large residuals produces a loose one.
This per-model adaptation is superior to global thresholds used by baselines. -
The system can be deployed using only ground truth pairs or none in the data-rich regime, as
validation-size insensitivity means a handful of ground-truth pairs (or none, in the data-rich regime where Jeffreys suffices) is enough to deploy near-optimally.
This drastically reduces the need for large, labeled validation sets. -
The system can distinguish between model selection and hyperparameter tuning by decoupling them:
a calibrates score values while the selected model is governed by the data, not the hyperparameter.
This prevents a single global threshold from being overly sensitive to miscalibration.
Abstract
The most widely used RANSAC variants score a candidate model by summing per-point scores that fall to zero beyond a threshold. These scores derive from the negative log-likelihood of a Gaussian-uniform mixture. The threshold separates inliers, which follow the noise model, from outliers, and it is a function of the observation noise level, also known as the inlier scale. We remove this dependence on the threshold by reversing the usual order of RANSAC inference: rather than estimating the scale and then scoring against it, we marginalize the scale in closed form for a fixed inlier partition and then optimize over partitions. The resulting score is maximized by a single sort followed by a linear sweep. A single closed-form expression spans informative and non-informative priors, so the same score adapts to datasets with extensive or limited ground-truth annotation. The proposed score has a single hyperparameter, the half-width of the uniform outlier distribution. We show that this half-width is far less sensitive to miscalibration than the inlier threshold and is easier to cross-validate. The proposed score is the first likelihood-based RANSAC score in which the inlier scale is absent from the formula. Project page: https://ransacscoringdoneright.github.io
Sources
- Progressive NAPSAC: sampling from gradually growing neighborhoods
- Fixing the RANSAC Stopping Criterion
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis