acceptodds
Withdrawn from ICLR 2027

DualU: Modeling Dual Uncertainties for Correspondence Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Correspondence learning aims to estimate geometric models from putative correspondences contaminated by severe outliers and is fundamental to many computer vision tasks. Existing methods have substantially improved outlier rejection by predicting inlier scores for each correspondence. However, under repeated structures or insufficient geometric context, outliers may exhibit patterns similar to inliers and consequently receive high scores, while even correctly identified inliers may still have different localization errors and therefore provide geometric measurements of varying accuracy. An inlier score alone does not quantify the reliability of the prediction or the geometric error of an inlier. To address this limitation, we propose DualU, a correspondence learning method that explicitly models dual uncertainties (i.e., the reliability of inlier predictions and the geometric errors of inliers), and uses them to guide geometric model estimation. Extensive experiments demonstrate that DualU outperforms state-of-the-art methods across multiple tasks and exhibits stronger generalization ability.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Correspondence learning, Uncertainty Estimation, Outlier Rejection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7ZqJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes DualU, a correspondence learning method that models two forms of uncertainty: the reliability of inlier predictions and the geometric residual variance of inliers. Built on DeMatch++, it combines an evidential inlier belief with predicted residual precision to weight correspondences in a differentiable geometric solver. Experiments evaluate relative pose estimation, homography estimation, visual localization, and generalization across scenes and feature descriptors.

Critical strengths

The paper addresses a reasonable limitation of probability-based correspondence weighting: correctly identified inliers can still differ substantially in geometric precision. The experimental evaluation covers several geometric estimation tasks and includes component ablations, a freely learned weighting baseline, and uncertainty quality analysis. The reported improvements over the shared DeMatch++ backbone are substantial in several settings, particularly outdoor pose and homography estimation, while introducing little additional computational overhead.

Critical weaknesses
  1. The necessity of dual uncertainty modeling is not fully established. In Table 6, residual uncertainty alone improves AUC@5° from 34.33 to 38.54, while the complete method reaches 39.45. Without variability estimates and more controlled analysis, the independent contribution of the inlier uncertainty branch remains unclear. 2. The evaluation lacks comparisons with representative dense matching pipelines. Consequently, the reported improvements over correspondence pruning methods do not establish the proposed method’s competitiveness within a broader geometric matching setting. 3. The paper also does not investigate integration with dense matching methods. It remains unclear whether DualU can improve correspondences produced by dense matchers.
Questions for authors
  1. Can you better isolate the contribution of the inlier uncertainty branch from the change in classification loss, and report whether its improvement over residual uncertainty alone is consistent across repeated runs? 2. How does DualU compare with representative dense matching pipelines ? 3. Have you attempted to apply DualU to correspondences obtained from dense matchers, with appropriate sampling where necessary? I would consider raising my score if the authors can convincingly address my concerns.
Reviewer_UQkXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DualU, a correspondence learning framework that explicitly models inlier uncertainty and residual uncertainty. These estimates are combined to guide weighted geometric estimation. Experiments on relative pose estimation, homography estimation, and visual localization demonstrate promising improvements over existing methods.

Critical strengths
  • The distinction between inlier prediction reliability and geometric measurement precision is well motivated.

  • The method achieves notable improvements across multiple benchmarks with negligible additional computational overhead.

Critical weaknesses
  • The effectiveness of inlier uncertainty remains insufficiently demonstrated. The authors motivate their approach by highlighting that outliers may receive high inlier scores under ambiguous geometric contexts. However, as acknowledged in Section 4.1, the positive belief mass remains close to the original inlier probability when the latter is high, limiting the direct effect of evidential uncertainty on such challenging correspondences. Although residual uncertainty may compensate for this limitation, the independent effectiveness of inlier uncertainty in addressing this motivation remains unclear.
  1. Limited methodological novelty and insufficient justification of the weighting formulation. The proposed framework primarily integrates established techniques, including evidential classification (Sensoy et al., NeurIPS 2018), heteroscedastic Gaussian uncertainty modeling (Kendall and Gal, NeurIPS 2017), and uncertainty-aware geometric estimation (Muhle et al., CVPR 2023). More directly, Cai et al. (ICLR 2025) have already explored evidential certainty estimation for feature matching. While the integration of these techniques into sparse correspondence learning is practically meaningful, the methodological novelty appears relatively limited.

Furthermore, Eq. 11 motivates a weight proportional to q_i / sigma_i^2, whereas Eq. 12 replaces q_i with the positive belief mass b_i^+. Although motivated by subjective logic, this substitution does not directly follow from the probabilistic derivation. While Table 6 demonstrates the benefit of introducing inlier uncertainty, it does not isolate the contribution of evidential training from that of positive belief weighting. A controlled comparison between q_i / sigma_i^2 and b_i^+ / sigma_i^2 under identical training objectives would better justify the proposed formulation.

Questions for authors
  • Can the proposed inlier uncertainty effectively identify and suppress high-confidence outliers, particularly in scenarios with repetitive structures or geometric ambiguities?

  • Could the authors clarify the key methodological differences between the proposed uncertainty formulation and existing approaches, particularly evidential classification (Sensoy et al., 2018; Cai et al., 2025) and uncertainty-aware geometric estimation (Muhle et al., 2023)? What is the main technical novelty beyond integrating these established techniques?

  • How does q_i/ sigma_i^2 compare with b_i^+/ sigma_i^2 under identical evidential training objectives?

Reviewer_viHgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DualU addresses correspondence learning for geometric estimation by distinguishing the reliability of an inlier prediction from the geometric precision of a true inlier. Built on DeMatch++, the method predicts an inlier probability qiq_i, a Beta distribution strength SiS_i, and a residual log-variance ℓi\ell_i. These outputs define the weight ρi=(qi−1/Si)exp⁡(−ℓi)\rho_i=(q_i-1/S_i)\exp(-\ell_i) for a differentiable geometric solver. Experiments cover relative pose estimation on YFCC100M and SUN3D, homography estimation on HPatches, and visual localization on Aachen Day-Night v1.1, with additional evaluations of descriptor transfer, component contributions, uncertainty estimates, and computational cost.

Critical strengths
  1. Meaningful gains across several geometric tasks. Compared with DeMatch++, DualU improves AUC@5° on unknown YFCC100M scenes from 34.33 to 39.45, HPatches accuracy at 3 pixels from 49.66 to 63.45, and Aachen nighttime localization at the intermediate threshold from 63.9 to 70.2. These results show benefits in both geometric estimation and downstream localization.
  2. Low overhead and encouraging transfer performance. Table 5 reports only about 0.010M additional parameters and 0.066 ms additional inference time. On YFCC100M, transfer to SuperPoint and XFeat improves AUC@20° over DeMatch++ by 8.97 and 11.84 percentage points, respectively. The component ablations and residual sparsification analysis provide further support for the method.
Critical weaknesses
  1. The independent contribution of inlier uncertainty is not sufficiently isolated. Equations (3)–(4) imply max⁡(0,2qi−1)≤bi+≤qi\max(0,2q_i-1)\leq b_i^+\leq q_i; at qi=0.99q_i=0.99, the evidence discount can only reduce the score to approximately 0.98. The paper acknowledges this limitation but does not establish how the strength branch helps identify confidently incorrect matches beyond changing qiq_i through joint training. Figure 3(b) also changes both the score and the final classification objective, from class-balanced BCE to IMSE plus KL, so its calibration improvement cannot be attributed to SiS_i alone. Table 6 shows only a 0.91-point AUC@5° improvement over the residual-only variant. Controls that separate these effects and repeated runs are needed to establish the contribution of the second uncertainty component.
  2. The interpretation and transfer of residual variance remain unclear. The learned variance concerns algebraic residuals, whose scale depends on coordinate and model normalization. Figure 3(a) supports residual ranking but does not establish variance calibration. More importantly, a variance trained on scalar epipolar residuals is reused for homography estimation, where each correspondence supplies two DLT constraints; how this output is applied and whether it retains an inverse-variance interpretation are insufficiently explained. The missing definition of Lgeo\mathcal L_{\mathrm{geo}}, unspecified variance safeguards, and unclear use of scores in HLoc also limit reproducibility. Since residual supervision excludes outliers, the possibility of falsely accepted matches receiving excessive precision needs examination.
Questions for authors
  1. Classification loss. Specify the supervision used in each Table 6 variant. Add a comparison that isolates the classification loss and includes a calibrated BCE baseline, reporting both calibration and pose accuracy.
  2. Numerical stability. Describe any variance floors or clipping, and test whether confidently misclassified outliers can receive excessive precision and dominate geometric fitting.
  3. Provide the full Lgeo\mathcal L_{\mathrm{geo}} definition and explain how scores enter HLoc/COLMAP. Clarify whether all baselines use the same filtering rules and solver settings.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.