Semi-Dense Matching Uncertainty Is Not Just Local Confidence

Hai Tran; Hoang-Tuan Nguyen; Huong Ninh; Khoa Hoang; Long Q. Tran

REVIEW 3 major objections 6 minor 46 references

Semi-Dense Matching Uncertainty Is Not Just Local Confidence

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read The paper claims that semi-dense matching uncertainty is not just local confidence: a 9-parameter post-hoc Laplace mixture that models coarse-assignment failures improves homography and pose estimation across six pretrained matchers.

desk verdict A useful plug-in for semi-dense matching uncertainty, with a real scope caveat the paper itself buries in the supplement. read the letter →

arxiv 2608.08685 v1 pith:XK2VS4VX submitted 2026-08-09 cs.CV

Khoa Hoang , Hoang-Tuan Nguyen , Huong Ninh , Hai Tran , Long Q. Tran This is my paper · ORCID

classification cs.CV

keywords semi-densematchinguncertaintyestimationcoarse-to-fineLaplacemixturemodelgeometricrefitrobusthomographyvisuallocalization

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

Coarse-to-fine semi-dense matchers produce two very different kinds of error: small local refinement noise and occasional catastrophic failures where the coarse search picks the wrong region entirely. The paper argues that standard uncertainty estimates, which only look at local refinement confidence, miss the second kind and are therefore overconfident on hard pairs. It proposes a post-hoc, calibrated two-component Laplace mixture, fit with only 9 parameters and no retraining, that captures both error regimes, and a refit step called CoRe that uses the posterior probability of coarse-assignment success as soft weights in the final geometric estimate. Across six pretrained matchers and five robust estimators, this consistently improves homography and pose accuracy at modest computational cost, especially on moderately hard pairs.

What carries the argument

The load-bearing object is the calibrated two-component Laplace mixture $\hat p_i(r) = (1-\alpha_i)\mathrm{Lap}(r|0,s_i^f) + \alpha_i\mathrm{Lap}(r|0,s_i^c)$, where $s_i^f$ and $s_i^c$ are per-axis scales calibrated from the matcher's fine- and coarse-level outputs and $\alpha_i = \sigma(w^\top\phi_i + \beta)$ is a sigmoid gate over coarse cues. Bayes' rule turns this likelihood, evaluated at the residual under an initial robust estimate $\theta_0$, into the coarse-success posterior $\omega_i$ used as a soft correspondence weight in one weighted geometric refit $\theta^\star = \arg\min_\theta \sum_i \omega_i \rho(x_i,\hat y_i;\theta)$. The mixture is what lets the method express both a sharp inlier peak and a heavy failure tail; the posterior refit is what converts that model into better geometry.

What would settle it

Take an image pair with a known homography, deliberately perturb the initial model by an increasing amount, and check whether the CoRe posterior still ranks matches by whether their coarse guess was correct; if the ranking degrades to chance as the perturbation grows past roughly 30 px, the weights are encoding geometry error rather than matching error, contradicting the central claim in that regime.

Watch

Extended reading notes

Core claim

The paper's central claim is that the overall error of a semi-dense coarse-to-fine matcher is a two-source mixture, and that current uncertainty estimates capture only one source. Concretely, it models the residual $\hat y_i - y_i$ as $(1-\alpha_i)\mathrm{Lap}(r|0,s_i^f) + \alpha_i\mathrm{Lap}(r|0,s_i^c)$: the fine component reflects local refinement noise when the coarse assignment is correct, and the coarse component absorbs the heavy tail of matches where the coarse search failed entirely. The paper then derives, via Bayes' rule, the posterior probability that the coarse assignment succeeded given the residual under an initial robust model, and uses that posterior as a soft weight in a single final geometric refit (CoRe). It reports that this refit improves homography AUC and pose accuracy across six pretrained matchers and five robust estimators, with parameters calibrated once on MegaDepth and transferred zero-shot.

Load-bearing premise

The refit assumes that residuals under the initial geometry estimate are close to the true matching errors, so when the initial geometry is wrong by tens of pixels the posterior weights stop measuring matching success; the paper's own supplementary analysis shows unstable behavior above roughly 30 px.

Editorial extensions

If this is right

Any coarse-to-fine matcher can improve downstream homography and pose accuracy by adding the calibrated mixture and CoRe, with no retraining and only tens of milliseconds of overhead.
Soft weighting with the coarse-success posterior beats hard inlier/outlier masking at 3px and looser thresholds, so standard RANSAC final steps discard useful residual information.
Fine-only uncertainty is bounded by the refinement window and cannot express large errors, so modeling the coarse-failure tail is necessary for calibrated uncertainty on hard pairs.
The 9 calibration parameters transfer zero-shot across datasets and resolutions, making calibration a one-time lightweight step.
Gains concentrate on moderately hard pairs with initial error between 1 and 30 pixels, while near-perfect fits may lose a little at the strict 1px threshold.

Reading between the lines

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

Beyond the paper's experiments, the same coarse-success posterior could be used inside iterative estimation loops: re-estimate the model, recompute residuals, recompute weights, and refit again, rather than only performing a single final refit.
Because the mixture is fit post-hoc from cues the matcher already outputs, the posterior could also serve as a soft inlier prior for training robust estimators end-to-end, which the paper does not explore.
The axis-factorized Laplace treats horizontal and vertical residuals independently; a correlated error model might behave differently on slanted or perspective-distorted surfaces, a testable variation not covered in the paper.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Punchline: the paper proposes a post-hoc, 9-parameter calibrated two-component Laplace mixture for semi-dense matcher uncertainty, and a coarse-success posterior refit (CoRe) that weights correspondences for geometric estimation. The core claim—that this improves homography and pose estimation across six frozen matchers and five robust estimators—holds in the aggregate, with gains that are often large and compute overhead that is genuinely small.

What is new: nobody else has offered a calibration stage that explicitly models coarse-assignment failures as a separate tail and then uses the posterior of coarse success as soft weights in a refit. The closest work, SURE, only models local uncertainty. The experiments are broad and the calibration is clean: fitted on MegaDepth, transferred zero-shot to HPatches, MTV, and Aachen. The ablation against uniform, matching-confidence, and binary weighting is convincing, and the support analysis shows the posterior tracks real coarse accuracy.

The soft spot is the one the stress-test identified. CoRe computes residuals under the initial geometry theta0 (Eq. 8). When theta0 is poor, the residuals are dominated by geometry error, so the posterior no longer measures coarse-assignment success. The paper's own Table 6 shows exactly this: for e0 < 1px, only 47.9% of pairs improve; for e0 > 30px, behavior is unstable. That means the abstract's 'consistently improves' is too strong. The accurate claim is 'improves reliably when the initial estimate is in a moderate-error band, roughly 1-30px, and is neutral or unstable outside it.' This is a real scope limitation and it belongs in the main text, not only the supplement. That said, the aggregate tables are not cherry-picked—most real pairs fall in the favorable band—so the method itself stands.

Minor issues: no error bars on the main tables; one qualitative MTV case in Fig. 3 reports a corner error of 3.7e14 px, which looks like a degenerate homography or a reporting bug; no comparison against SURE; and Eq. (3) assumes all coarse failures collapse into one component, which is a practical approximation but deserves one sentence of justification.

Who this is for: anyone building geometric estimation pipelines with semi-dense matchers, and anyone working on uncertainty calibration for correspondence. It deserves a serious referee. I would accept it for review and ask for the scope claim to be fixed, the broken example to be repaired, and a small sensitivity experiment in the main paper.

Referee Report

3 major / 6 minor

Summary. The paper addresses uncertainty quantification for coarse-to-fine semi-dense image matchers. It proposes a post-hoc calibrated two-component Laplace mixture model with nine learnable parameters: one component captures local refinement noise and the other captures the heavier tail of coarse-assignment failures. A learned gate combines coarse and fine uncertainty cues into a mixture likelihood. The authors then introduce CoRe (Coarse-success posterior Refit), a final geometric refitting step that uses the posterior probability of coarse-assignment success, computed under an initial robust estimate, as soft weights over all correspondences. The method is evaluated on homography estimation (HPatches, MTV) and visual localization (Aachen v1.1) across six pretrained matchers and five robust estimators, with parameters calibrated on MegaDepth and transferred zero-shot. The main claim is that CoRe consistently improves downstream geometric accuracy with minimal computational overhead.

Significance. If the claims hold, the contribution is practically valuable: it requires no retraining, adds only nine calibration parameters, applies to multiple off-the-shelf matchers and robust estimators, and is accompanied by released code and a broad experimental study. The paper also contains useful ablations comparing different weighting strategies, robustness across estimators, and a supplementary sensitivity analysis that is more informative than is typical. The external evaluation on datasets not used for calibration (HPatches, MTV, Aachen) and the explicit zero-shot transfer are strengths, as is the careful distinction between fine-only uncertainty and the proposed mixture. However, the central 'consistent improvement' claim is stronger than the evidence supports, particularly with respect to dependence on the initial geometric estimate and to several reported decreases in Table 2 and Table 4. The method is defensible as a lightweight way to improve geometry in the moderate-error regime, but the paper needs to scope the claim and address the failure regimes it already documents.

major comments (3)

[§3.3 (Eqs. 8–10); Supp B.1 (Table 6, Fig. 7)] The claim that CoRe 'consistently improves' geometric estimation is not supported over the full range of initial geometry quality. The coarse-success posterior in Eq. (10) is evaluated on the reprojection residual r^(0)_i = yhat_i − Π_{θ0}(x_i), so when the initial θ0 is inaccurate, the residual is dominated by geometry error rather than by matching error; correct correspondences can then be down-weighted, exactly when they are needed to correct θ0. The paper's own Table 6 shows that for e0 < 1 px, CoRe improves only 47.9% of pairs with a median change of −0.004 px, and for e0 > 30 px the improvement frequency is 58.3–100% in bins containing only 3–12 pairs, i.e., unstable. Figure 7 confirms the same pattern across matchers. The authors should report the distribution of initial errors e0 in the evaluation datasets, condition the headline results on e0 or on difficulty, and either add a safeguard that retains the initial model when the posterior is unreliable or revise the universal improvement claim.
[Abstract and Tables 1–4] The abstract's statement that the method 'consistently improves downstream geometric accuracy across various pretrained-only matchers and robust estimators' is contradicted by several reported cells. In Table 2, GIM-LoFTR Night at (0.25 m, 2°) decreases from 68.6 to 67.0, EfficientLoFTR Night decreases from 73.3 to 72.8, and CoMatch Day decreases from 85.2 to 85.1. In Table 4, the 1 px AUC decreases for LO-RANSAC, PROSAC, and GC-RANSAC. The main text correctly says 'in most settings' for localization, but the abstract and contribution (iii) should be reworded to reflect improvements at moderate thresholds and difficulty levels rather than universal consistency, or the authors should provide error bars and statistical tests supporting the aggregate claim.
[§3.2, Eq. (3)] Equation (3) collapses the marginalization over all incorrect coarse cells into a single coarse-error component p_c_i(r). This is a strong approximation: the residual distribution after a wrong coarse assignment depends on which coarse cell was selected, so the two-component mixture is not an exact generative model. Consequently, the weight ω_i in Eq. (10) should be described as a calibrated heuristic rather than an exact posterior probability of coarse-assignment success. Because CoRe's usefulness depends on this posterior tracking true coarse success, the current validation in Supp Fig. 8 is only aggregate and should be augmented by conditioning on the initial error e0; otherwise the failure regime identified above also undermines the posterior interpretation.

minor comments (6)

[Fig. 3] The corner error value '372445508502019.8 -> 296482922247990.4 px' appears to be a typo or numerical overflow; please correct it.
[Algorithm 1 and Sec. 4.1] Calibration uses ground-truth correspondences y_i on MegaDepth, but the paper does not state how these correspondences are computed (e.g., depth and pose warping, pseudo-ground-truth, or manual labeling); this information is needed for reproducibility.
[Eq. (5)] The square-root notation over the learned multipliers a and b is confusing; please clarify whether the learnable parameters are the spatial scales themselves or their squares.
[Supp. A and Sec. 3.2] The mapping from the main-text notation Θ = {a, b, w, β} to the nine parameters {a_x, a_y, b_x, b_y, k_s, k_m, t_x, t_y, t_m} should appear in the main paper, since the main text never defines the dimensions of w and β.
[Table 3] The baseline weighting strategies 'mconf refit' and 'Raw fine std' should be defined in the caption or in the text, because their exact construction is not otherwise specified.
[Sec. 3.1 and Abstract] There are minor text issues: 'solely only captures' is redundant, and the code URL in the abstract contains an unwanted line break ('Probabilistic\n-matching').

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: calibration is external, downstream evaluation is zero-shot, and CoRe's residual-based weighting is a standard application of the fitted model.

full rationale

The paper's central derivation is self-contained. The 9 calibration parameters {a,b,w,beta} are fitted by minimizing negative log-likelihood on a held-out MegaDepth calibration set (Eq. 7, Algorithm 1), and the fitted model is then transferred without dataset-specific tuning to HPatches, MTV, and Aachen v1.1 (Sec. 4.1, Tables 1-2). The downstream benchmarks are external and unused during fitting, so the reported improvements are genuine out-of-sample predictions rather than re-statements of fit residuals. CoRe computes coarse-success posteriors from reprojection residuals under an initial robust estimate (Eqs. 8-10) and uses them as soft weights in a final refit (Eq. 11); this is standard calibrated-model usage, not a quantity defined to reproduce its own input. The paper's own supplementary analysis (Supp B.1, Eq. 14, Table 6) explicitly discloses that CoRe becomes less stable when the initial geometry error exceeds 30 px and that easy pairs with e0 < 1 px show near-neutral behavior. This is a scope limitation, and the paper flags it; it is not circularity. No self-citation is load-bearing: the cited prior matchers and robust estimators are external baselines, and there is no invoked uniqueness theorem or ansatz smuggled in from the authors' own prior work. Consequently, the derivation chain does not reduce to its inputs at any step.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on 9 fitted calibration parameters (a, b, k_s, k_m, t_x, t_y, t_m) and on domain assumptions about error decomposition and initial geometry quality. No new physical or conceptual entities are introduced.

free parameters (5)

Calibrated coarse scale multiplier a (ax, ay) = Learned on MegaDepth calibration set; numeric values not reported
Maps the raw coarse distribution spread to pixel-error scale by sc_i = sqrt(a) * tilde_sc_i (Eq. 5).
Calibrated fine scale multiplier b (bx, by) = Learned on MegaDepth calibration set; numeric values not reported
Maps raw fine refinement scale to pixel-error scale by sf_i = sqrt(b) * tilde_sf_i (Eq. 5).
Gate weight ks for normalized coarse scale = Learned on MegaDepth; numeric values not reported
Controls how the image-size-normalized coarse scale shifts the coarse-failure prior in Eq. (6); see Supp A for the axis-wise sigmoid form.
Gate weight km for negative log matching confidence = Learned on MegaDepth; numeric values not reported
Controls how low matching confidence increases the coarse-failure prior.
Gate thresholds tx, ty, tm = Learned on MegaDepth; numeric values not reported
Provide the offsets in the sigmoid gate for the x-axis, y-axis, and confidence terms; these are part of the 9 calibration parameters.

assumptions (4)

domain assumption Conditioned on a wrong coarse assignment, the residual distribution can be approximated by a single coarse-error component p_c_i(r) that does not depend on the specific wrong cell (Eq. 3).
This approximation collapses the sum over all wrong coarse cells into one component; it is essential for the two-component mixture to be tractable.
domain assumption The initial robust estimate theta_0 is accurate enough that residuals r_i^(0) approximate the true matching errors r_i (Eq. 8).
CoRe weights are computed from r_i^(0); if the initial geometry is poor, the posterior weights are distorted. The paper analyzes this only in Supp B.1.
standard math Bayes' theorem is applied to derive the coarse-success posterior in Eq. (9).
Standard probability rule, used to convert the fitted mixture likelihood and prior into the posterior weight.
domain assumption Raw fine and coarse uncertainty cues from pretrained matchers are informative for the true spatial error, up to a per-axis multiplicative calibration (Eq. 5).
The calibration relies on these cues; if they carry no signal, the 9-parameter fit cannot recover it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Semi-Dense Matching Uncertainty Is Not Just Local Confidence." pith.science (2026). https://pith.science/paper/XK2VS4VX

@misc{pith2026260808685,
  author       = {Pith},
  title        = {Pith review of: Semi-Dense Matching Uncertainty Is Not Just Local Confidence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XK2VS4VX}},
  note         = {Machine review of arXiv:2608.08685}
}

read the original abstract

Reliable semi-dense matching is essential for modern geometric vision systems. Designed under a coarse-to-fine paradigm, it achieves an optimal balance between performance and computational cost. However, existing methods often struggle to provide well-quantified uncertainties, where catastrophic coarse-assignment failures are ignored, leading to truncated error distributions and severely misjudged geometric estimations. In this paper, we propose a lightweight, post-hoc overall uncertainty estimation framework that introduces a two-component calibrated Laplace mixture model with only 9 learnable parameters. The objective is to explicitly capture both the sharp local refinement noise and the broader tail of coarse-assignment failures. We introduce the Coarse-success posterior Refit (CoRe) method, a geometric refitting module that utilizes the posterior probability of coarse-assignment success as soft correspondence weights. Extensive experiments show that our method consistently improves downstream geometric accuracy across various pretrained-only matchers and robust estimators with minimal computational overhead. Our code is available at https://github.com/khoavpt/Probabilistic-matching.

Figures

Figures reproduced from arXiv: 2608.08685 by the authors.

**Figure 1.** Coarse-success posterior refit improves downstream homography estimation. Radar plots compare HPatches AUC@3px before and after our refit. The left panel varies the matcher with RANSAC fixed, while the right varies the robust estimator with ELoFTR fixed. However, coarse-to-fine matchers frequently struggle to provide well-quantified uncertainties for their predictions. This two-stage design fundamentally fractures… view at source ↗

**Figure 2.** Overview of the proposed framework. Left: the conventional coarse-to-fine matching and robust geometry estimation pipeline. Right: our framework calibrates coarse- and fine-level cues into a two-component error likelihood (Alg. 1), which is used to estimate coarse-assignment success posteriors and perform our posterior-weighted geometric refit (CoRe; Alg. 2). assignment, frequently failing to properly evaluate over… view at source ↗

**Figure 3.** Qualitative homography refit. This figure shows the effect of applying CoRe to homography estimation. Rows show increasingly difficult pairs from top to bottom, and columns compare stronger matchers from left to right. Matches are colored by Expected abs. error scale (px), from low uncertainty in green to high uncertainty in red, with gray matches de-emphasized. CoRe is applied on all matches including gray outlier… view at source ↗

Figures from the paper (8 more)

**Figure 4.** Figure 4: Tail probability assignment. The model dynamically assigns higher tail mass (P(|rx| > 8px)) on harder datasets and for weaker matchers, demonstrating adaptive calibration. existing coarse-to-fine matchers [21, 23, 37, 44]. Since these matchers only run the fine refinem…

**Figure 5.** Figure 5: Resolution transfer of the calibrated error model. The model is calibrated once on 200 held-out MegaDepth pairs at 960×640 and evaluated without retuning on MegaDepth-1500 at different resolutions. Top: signed residual distribution for one representative coordinate axi…

**Figure 7.** Figure 7: Sensitivity to initial geometry across matchers. Top: median relative ACE reduction after CoRe across ranges of initial ACE e0. Bottom: fraction of pairs improved by CoRe. The marker area reflects the number of valid pairs in each bin. −40% −20% 0% 20% 40% 60% 100(1 − …

**Figure 8.** Figure 8: shows that our method correctly assigns higher posterior values to more accurate coarse assignments. The predicted probabilities are not perfectly aligned with the empirical success frequencies, which is expected because the uncertainty model is calibrated on MegaDepth…

**Figure 9.** Figure 9: Impact of calibration data quantity on convergence and computational efficiency. (a) The calibration objective function (Negative Log-Likelihood) over varying numbers of calibration pairs. (b) Optimization time, which scales almost linearly with the number of samples…

**Figure 10.** Figure 10: Impact of calibration set size on downstream performance and generalization. We evaluate the HPatches AUC@3px (a) and mean per-axis NLL on MegaDepth1500 (b) as functions of the number of calibration pairs (N). Our calibration algorithm demonstrates exceptional sampl…

**Figure 11.** Figure 11: Uncertainty under controlled directional blur. Predicted spatial error scales are visualized as ellipses around the matches. Horizontal and vertical Gaussian blur increase the estimated uncertainty primarily along the corresponding blur direction. Furthermore, due …

**Figure 12.** Figure 12: Additional qualitative CoRe refinement results. We compare the initial and CoRe-refined homographies across four datasets and three representative matchers. CoRe improves alignment in most examples, while some challenging MTV cases illustrate representative failure mo…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 39 canonical work pages

[1]

Netvlad: CNN architecture for weakly supervised place recognition

Relja Arandjelovic, Petr Gron ´at, Akihiko Torii, Tom ´as Pa- jdla, and Josef Sivic. Netvlad: CNN architecture for weakly supervised place recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV , USA, June 27-30, 2016, pages 5297–5307. IEEE Computer Society, 2016. 6

work page 2016
[2]

Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors

Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krys- tian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 3852–3861. IEEE Computer Society, 2017. 6, 7, 8, 12, 13

work page 2017
[3]

Graph-cut RANSAC

Daniel Barath and Jiri Matas. Graph-cut RANSAC. In2018 IEEE Conference on Computer Vision and Pattern Recog- nition, CVPR 2018, Salt Lake City, UT, USA, June 18- 22, 2018, pages 6733–6741. Computer Vision Foundation / IEEE Computer Society, 2018. 2, 5, 6, 7, 8

work page 2018
[4]

MAGSAC: marginalizing sample consensus

Daniel Barath, Jiri Matas, and Jana Noskova. MAGSAC: marginalizing sample consensus. InIEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 10197– 10205. Computer Vision Foundation / IEEE, 2019. 2

work page 2019
[5]

Magsac++, a fast, reliable and accurate robust estima- tor

D ´aniel Bar´ath, Jana Noskova, Maksym Ivashechkin, and Jiri Matas. Magsac++, a fast, reliable and accurate robust estima- tor. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 1301–1309. Computer Vision Founda- tion / IEEE, 2020. 2, 5, 6, 7, 8

work page 2020
[6]

Aspanformer: Detector-free image matching with adaptive span transformer

Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming- min Zhen, Tian Fang, David Mckinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. InEuropean conference on computer vision, pages 20–36. Springer, 2022. 2

work page 2022
[7]

Matching with PROSAC - pro- gressive sample consensus

Ondrej Chum and Jiri Matas. Matching with PROSAC - pro- gressive sample consensus. In2005 IEEE Computer Soci- ety Conference on Computer Vision and Pattern Recognition (CVPR 2005), 20-26 June 2005, San Diego, CA, USA, pages 220–226. IEEE Computer Society, 2005. 2, 5, 6, 7, 8

work page 2005
[8]

Locally opti- mized RANSAC

Ondrej Chum, Jiri Matas, and Josef Kittler. Locally opti- mized RANSAC. InPattern Recognition, 25th DAGM Sym- posium, Magdeburg, Germany, September 10-12, 2003, Pro- ceedings, pages 236–243. Springer, 2003. 5, 6, 7, 8

work page 2003

Show all 46 references

[9]

Dkm: Dense kernelized feature matching for geometry estimation

Johan Edstedt, Ioannis Athanasiadis, M ˚arten Wadenb ¨ack, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17765–17775, 2023. 2

2023
[10]

RoMa v2: Harder Better Faster Denser Feature Match- ing.arXiv preprint arXiv:2511.15706, 2025

Johan Edstedt, David Nordstr ¨om, Yushan Zhang, Georg B¨okman, Jonathan Astermark, Viktor Larsson, Anders Hey- den, Fredrik Kahl, M ˚arten Wadenb¨ack, and Michael Fels- berg. RoMa v2: Harder Better Faster Denser Feature Match- ing.arXiv preprint arXiv:2511.15706, 2025. 2, 3

2025 arXiv
[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.Commun. ACM, 24(6):381–395, 1981. 2, 5, 6, 7, 8

1981
[12]

Dropout as a bayesian approximation: Representing model uncertainty in deep learning

Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. Ininternational conference on machine learning, pages 1050–1059. PMLR, 2016. 2

2016
[13]

Top- icfm: Robust and interpretable topic-assisted feature match- ing

Khang Truong Giang, Soohwan Song, and Sungho Jo. Top- icfm: Robust and interpretable topic-assisted feature match- ing. InProceedings of the AAAI conference on artificial in- telligence, pages 2447–2455, 2023. 2

2023
[14]

Strictly proper scor- ing rules, prediction, and estimation.Journal of the Ameri- can Statistical Association, 102(477):359–378, 2007

Tilmann Gneiting and Adrian E Raftery. Strictly proper scor- ing rules, prediction, and estimation.Journal of the Ameri- can Statistical Association, 102(477):359–378, 2007. 5

2007
[15]

Weinberger

Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. InProceedings of the 34th International Conference on Machine Learning - Volume 70, page 1321–1330. JMLR.org, 2017. 3

2017
[16]

Cambridge University Press, 2 edition, 2003

Richard Hartley and Andrew Zisserman.Multiple View Ge- ometry in Computer Vision. Cambridge University Press, 2 edition, 2003. 2, 3, 5

2003
[17]

An adaptive confidence measure for optical flows based on linear subspace projections

Claudia Kondermann, Daniel Kondermann, Bernd J ¨ahne, and Christoph Garbe. An adaptive confidence measure for optical flows based on linear subspace projections. InJoint Pattern Recognition Symposium, pages 132–141. Springer,
[18]

A statistical confidence measure for optical flows

Claudia Kondermann, Rudolf Mester, and Christoph Garbe. A statistical confidence measure for optical flows. InEu- ropean Conference on Computer Vision, pages 290–301. Springer, 2008. 2

2008
[19]

Accurate uncertainties for deep learning using calibrated re- gression

V olodymyr Kuleshov, Nathan Fenner, and Stefano Ermon. Accurate uncertainties for deep learning using calibrated re- gression. InProceedings of the 35th International Confer- ence on Machine Learning, ICML 2018, Stockholmsm¨assan, Stockholm, Sweden, July 10-15, 2018, pages 280...

2018
[20]

SURE: semi-dense uncertainty- refined feature matching.CoRR, abs/2603.04869, 2026

Sicheng Li, Zaiwang Gu, Jie Zhang, Qing Guo, Xudong Jiang, and Jun Cheng. SURE: semi-dense uncertainty- refined feature matching.CoRR, abs/2603.04869, 2026. 1, 2

2026
[21]

EDM: efficient deep fea- ture matching

Xi Li, Tong Rao, and Cihui Pan. EDM: efficient deep fea- ture matching. InIEEE/CVF International Conference on Computer Vision, ICCV 2025, Honolulu, HI, USA, October 19-25, 2025, pages 26198–26208. IEEE, 2025. 1, 2, 3, 4, 5, 6, 11

2025
[22]

Megadepth: Learning single- view depth prediction from internet photos

Zhengqi Li and Noah Snavely. Megadepth: Learning single- view depth prediction from internet photos. InComputer Vision and Pattern Recognition (CVPR), 2018. 6, 11, 12, 13

2018
[23]

Comatch: Dynamic covisibility-aware transformer for bilateral subpixel-level semi-dense image matching

Zizhuo Li, Yifan Lu, Linfeng Tang, Shihua Zhang, and Ji- ayi Ma. Comatch: Dynamic covisibility-aware transformer for bilateral subpixel-level semi-dense image matching. In IEEE/CVF International Conference on Computer Vision, ICCV 2025, Honolulu, HI, USA, October 19-25, 2025, ...

2025
[24]

A multi-view thermal- visible image dataset for cross-spectral matching.Remote

Yuxiang Liu, Yu Liu, Shen Yan, Chen Chen, Jikun Zhong, Yang Peng, and Maojun Zhang. A multi-view thermal- visible image dataset for cross-spectral matching.Remote. Sens., 15(1):174, 2023. 6, 13

2023
[25]

Learning a confidence measure for optical 9 flow.IEEE transactions on pattern analysis and machine intelligence, 35(5):1107–1120, 2012

Oisin Mac Aodha, Ahmad Humayun, Marc Pollefeys, and Gabriel J Brostow. Learning a confidence measure for optical 9 flow.IEEE transactions on pattern analysis and machine intelligence, 35(5):1107–1120, 2012. 2

2012
[26]

Orb-slam2: An open- source slam system for monocular, stereo, and rgb-d cam- eras.IEEE transactions on robotics, 33(5):1255–1262, 2017

Raul Mur-Artal and Juan D Tard ´os. Orb-slam2: An open- source slam system for monocular, stereo, and rgb-d cam- eras.IEEE transactions on robotics, 33(5):1255–1262, 2017. 1

2017
[27]

Orb-slam: A versatile and accurate monocular slam system.IEEE transactions on robotics, 31(5):1147–1163,

Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: A versatile and accurate monocular slam system.IEEE transactions on robotics, 31(5):1147–1163,
[28]

Potje, Felipe Cadar, Andr ´e Ara ´ujo, Renato Martins, and Erickson R

Guilherme A. Potje, Felipe Cadar, Andr ´e Ara ´ujo, Renato Martins, and Erickson R. Nascimento. Xfeat: Accelerated features for lightweight image matching. InIEEE/CVF Con- ference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 2682–

2024
[29]

An efficient method for finding the min- imum of a function of several variables without calculating derivatives.The computer journal, 7(2):155–162, 1964

Michael JD Powell. An efficient method for finding the min- imum of a function of several variables without calculating derivatives.The computer journal, 7(2):155–162, 1964. 6

1964
[30]

From coarse to fine: Robust hierarchical localization at large scale

Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12716–12725, 2019. 1, 6

2019
[31]

Image retrieval for image-based localization revis- ited

Torsten Sattler, Tobias Weyand, Bastian Leibe, and Leif Kobbelt. Image retrieval for image-based localization revis- ited. InBMVC, page 4, 2012. 1

2012
[32]

Benchmarking 6dof outdoor visual local- ization in changing conditions

Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, Fredrik Kahl, and Tom´as Pajdla. Benchmarking 6dof outdoor visual local- ization in changing conditions. In2018 IEEE Confer...

2018
[33]

Structure- from-motion revisited

Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 4104–4113, 2016. 1

2016
[34]

Ransac scoring functions: Analysis and reality check.arXiv preprint arXiv:2512.19850, 2025

Alexander Shekhovtsov. Ransac scoring functions: Analysis and reality check.arXiv preprint arXiv:2512.19850, 2025. 3

2025
[35]

Gim: Learning generalizable image matcher from internet videos

Xuelun Shen, Zhipeng Cai, Wei Yin, Matthias M ¨uller, Zijun Li, Kaixuan Wang, Xiaozhi Chen, and Cheng Wang. Gim: Learning generalizable image matcher from internet videos. InThe Twelfth International Conference on Learning Repre- sentations, 2024. 6, 11

2024
[36]

Hao Song, Tom Diethe, Meelis Kull, and Peter A. Flach. Dis- tribution calibration for regression. InProceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, pages 5897–5906. PMLR, 2019. 3, 5

2019
[37]

LoFTR: Detector-free local feature matching with transformers.CVPR, 2021

Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers.CVPR, 2021. 1, 2, 3, 4, 5, 6, 11

2021
[38]

Quadtree attention for vision transformers.arXiv preprint arXiv:2201.02767, 2022

Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers.arXiv preprint arXiv:2201.02767, 2022. 2

2022 arXiv
[39]

Learning accurate dense correspondences and when to trust them

Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5714–5724,
[40]

Pdc-net+: Enhanced probabilistic dense corre- spondence network.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10247–10266, 2023

Prune Truong, Martin Danelljan, Radu Timofte, and Luc Van Gool. Pdc-net+: Enhanced probabilistic dense corre- spondence network.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10247–10266, 2023. 2

2023
[41]

Attention is all you need.Advances in neural information processing systems, 30, 2017

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 2

2017
[42]

Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, St ´efan J

Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, St ´efan J. van der Walt, Matthew Brett, Joshua Wil- son, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelso...

2020
[43]

Matchformer: Interleaving attention in transformers for feature matching

Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. InProceedings of the Asian conference on computer vision, pages 2746–2762,
[44]

Efficient LoFTR: Semi-dense local feature matching with sparse-like speed

Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient LoFTR: Semi-dense local feature matching with sparse-like speed. InCVPR, 2024. 1, 3, 4, 6, 7, 11 10 Semi-Dense Matching Uncertainty Is Not Just Local Confidence Supplementary Material This supplementary ...

2024
[2691]

1, 6, 11

IEEE, 2024. 1, 6, 11

2024
[8610]

Computer Vision Foundation / IEEE Computer Soci- ety, 2018. 6

2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Netvlad: CNN architecture for weakly supervised place recognition

Relja Arandjelovic, Petr Gron ´at, Akihiko Torii, Tom ´as Pa- jdla, and Josef Sivic. Netvlad: CNN architecture for weakly supervised place recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV , USA, June 27-30, 2016, pages 5297–5307. IEEE Computer Society, 2016. 6

work page 2016

[2] [2]

Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors

Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krys- tian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 3852–3861. IEEE Computer Society, 2017. 6, 7, 8, 12, 13

work page 2017

[3] [3]

Graph-cut RANSAC

Daniel Barath and Jiri Matas. Graph-cut RANSAC. In2018 IEEE Conference on Computer Vision and Pattern Recog- nition, CVPR 2018, Salt Lake City, UT, USA, June 18- 22, 2018, pages 6733–6741. Computer Vision Foundation / IEEE Computer Society, 2018. 2, 5, 6, 7, 8

work page 2018

[4] [4]

MAGSAC: marginalizing sample consensus

Daniel Barath, Jiri Matas, and Jana Noskova. MAGSAC: marginalizing sample consensus. InIEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 10197– 10205. Computer Vision Foundation / IEEE, 2019. 2

work page 2019

[5] [5]

Magsac++, a fast, reliable and accurate robust estima- tor

D ´aniel Bar´ath, Jana Noskova, Maksym Ivashechkin, and Jiri Matas. Magsac++, a fast, reliable and accurate robust estima- tor. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 1301–1309. Computer Vision Founda- tion / IEEE, 2020. 2, 5, 6, 7, 8

work page 2020

[6] [6]

Aspanformer: Detector-free image matching with adaptive span transformer

Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming- min Zhen, Tian Fang, David Mckinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. InEuropean conference on computer vision, pages 20–36. Springer, 2022. 2

work page 2022

[7] [7]

Matching with PROSAC - pro- gressive sample consensus

Ondrej Chum and Jiri Matas. Matching with PROSAC - pro- gressive sample consensus. In2005 IEEE Computer Soci- ety Conference on Computer Vision and Pattern Recognition (CVPR 2005), 20-26 June 2005, San Diego, CA, USA, pages 220–226. IEEE Computer Society, 2005. 2, 5, 6, 7, 8

work page 2005

[8] [8]

Locally opti- mized RANSAC

Ondrej Chum, Jiri Matas, and Josef Kittler. Locally opti- mized RANSAC. InPattern Recognition, 25th DAGM Sym- posium, Magdeburg, Germany, September 10-12, 2003, Pro- ceedings, pages 236–243. Springer, 2003. 5, 6, 7, 8

work page 2003