acceptodds
Under review as a conference paper at ICLR 2027

Beyond Inlier Confidence: Learning Residual Quantiles for Correspondence Pruning

Abstract

Learned two-view correspondence pruning commonly uses scalar inlier scores to guide geometric estimation. However, correspondences with similar probabilities of satisfying an inlier threshold can have different residual distributions. We propose Residual Quantile Regression (RQR), which augments inlier classification with conditional geometric residual quantiles. Using a pinball objective, RQR learns ordered quantiles of the log epipolar residual for retained correspondences, providing a compact description of residual magnitude and spread. These quantiles guide correspondence learning during training; in the inference-time realization, RQR-F, they condition features through gated updates, while their spreads inform weights for differentiable weighted eight-point estimation. This formulation makes residual statistics explicit in correspondence representations and geometric weighting. Experiments on relative pose and homography estimation demonstrate the effectiveness of residual quantile learning across pruning architectures and geometric tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Matching, Correspondence Learning, Pose Estimation, Quantile Regression

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KnXDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes Residual Quantile Regression (RQR) for learned two-view correspondence pruning. It augments thresholded inlier classification with pinball-supervised quantiles of transformed epipolar residuals. Two training-only integrations use these predictions either to guide native correspondence scores or to supervise shared features. An inference-time variant, RQR-F, retains the quantile descriptor to condition features and modify weights for weighted eight-point estimation. Evaluation includes paired comparisons across six pruning architectures, relative-pose estimation, homography transfer, cross-domain and cross-feature testing, and a fixed-host RQR-F ablation. The results demonstrate benefits in several settings, while also showing substantial dependence on the host architecture, supervision placement, and geometric estimator.

Critical strengths

The paper identifies a meaningful limitation of thresholded inlier labels and shows that additional residual supervision can benefit an existing pruning pipeline. In particular, BCLNet’s reported weighted AUC@5 improves from 33.36 to 37.13. Since the auxiliary prediction heads are removed at inference, this improvement comes without additional inference modules. This is a useful practical result, although it primarily supports the value of residual supervision rather than establishing the specific advantage of multiple quantiles.

Critical weaknesses
  1. The quantile-specific contribution is not established. In Table 4, Huber outperforms three-quantile RQR under the matched all-interface comparison, while the stronger final-interface RQR result lacks corresponding Huber/Laplace controls. Table 3 also bundles quantile prediction, supervision, and weighting. These experiments do not isolate the benefit of multiple quantiles from continuous residual supervision, supervision placement, or additional weighting.

  2. The smaller improvements lack repeatability support. Each configuration uses a single training run. The paired bootstrap measures test-pair variability, not training variability, leaving the modest gains on several hosts and the incremental prompting benefit insufficiently substantiated.

  3. The presentation is substantially below the standard expected of a research paper. The method section is unnecessarily opaque: key definitions are delayed or incomplete, and three materially different implementations are interwoven without a coherent account of their data flow, objectives, and gradient paths. The prose repeatedly relies on abstract, formulaic descriptions without providing the concrete explanations needed to understand the computation. Readers must reconstruct the actual pipelines by piecing together scattered passages across the main text and appendices.

Questions for authors
  1. Quantile-specific contribution: How does final-interface RQR compare with Huber and Laplace supervision under matched training schedules and validation-selection budgets? For RQR-F, how does quantile supervision alone compare with adding the proposed inference-time weighting? These comparisons would clarify whether multiple quantiles provide benefits beyond continuous residual supervision and its placement.

  2. Repeatability: How consistent are the reported improvements across independent training runs? Please report means and variability for representative host comparisons and the RQR-F ablation, particularly whether the smaller gains and the incremental prompting benefit persist across seeds.

  3. Method specification: Can the authors provide a consolidated algorithm or diagram for each of the three integrations, explicitly identifying the residual transformation, prediction heads, loss terms, gradient paths—including detached connections—and components retained at inference? Please also resolve the delayed or incomplete symbol definitions so that each pipeline can be understood and reconstructed unambiguously.

6 comments
AuthorsOpenReview ↗

K1. Quantile targets at the final prediction stage

Thank you for identifying supervision placement as a confound in the original residual-target comparison.

Concern. How does final-stage RQR compare with Huber and Laplace when the training schedule and validation-selection budget are matched?

Response. RQR has higher weighted AUC@5 than both matched final-stage controls, with positive paired intervals for both differences. The final-stage comparison uses the same BCLNet architecture, supervision location, three-output head capacity, initialization, classification and geometric objectives, and residual-loss coefficient (0.25). The objectives share the same training schedule and validation-selection budget. All auxiliary heads are removed at inference.

Table 4 and Appendix Table B.4 report the following pose AUCs (%):

Final-stage objectiveWeighted AUC@5 / 10 / 20RANSAC AUC@5 / 10 / 20
Huber35.78 / 56.75 / 73.3734.82 / 54.99 / 71.88
Laplace36.19 / 57.24 / 73.8134.91 / 55.14 / 72.26
RQR, three quantiles37.13 / 57.80 / 73.8135.19 / 55.14 / 71.99

Appendix B.3 quantifies the weighted differences in AUC points:

Final-stage comparisonWeighted AUC@5 gain [95% CI]Weighted AUC@10 gain [95% CI]
RQR minus Huber+1.35 [+0.60, +2.12]+1.05 [+0.42, +1.71]
RQR minus Laplace+0.93 [+0.14, +1.72]+0.55 [-0.13, +1.23]

This comparison changes the residual objective while holding its location and capacity fixed. Huber applies Smooth L1 to three predictions of the same scalar target; Laplace learns a location and positive scale; RQR supervises three ordered levels. Thus the result supports a benefit of the quantile objective beyond adding continuous supervision at this location. It does not identify quantile asymmetry alone as the cause.

The all-stage results in Table 4 remain: Huber and Laplace outperform three-quantile RQR there. At the final stage, RQR improves tight-angle weighted pose accuracy; at 20 degrees it matches Laplace to the reported precision. The corresponding RANSAC difference intervals include zero. The evidence therefore ties the quantile advantage to supervision placement and the geometric estimator.

Revision. Section 4.4 and Table 4 (p. 9) now separate all-stage and final-stage supervision. Appendix B.3/Table B.4 (p. 16) specifies the matched objectives and intervals; the conclusion (p. 10) limits the advantage to tighter weighted-pose tolerances.

AuthorsOpenReview ↗

K2. Separating residual supervision from inference weighting

Thank you for pointing out that the original ablation combined quantile supervision and weighting.

Concern. Does quantile supervision alone help, and what does using the predicted gaps at inference contribute beyond that supervision?

Response. We separate the trained supervision-only configuration from an intervention on the weights of the same trained model. The latter gives the direct test of whether predicted gaps are useful during geometric estimation.

In Table B.2, each Off/On pair shares identical trained parameters and differs only in the explicit gap penalties:

Feature promptGap weighting at inferenceWeighted AUC@5 / 10 / 20
NoOff38.13 / 59.50 / 75.32
NoOn38.90 / 59.98 / 75.63
YesOff38.55 / 59.73 / 75.74
YesOn39.15 / 60.12 / 75.94

Using the gaps improves weighted AUC at all three tolerances with either prompt configuration. The paired weighted-gain intervals are positive (Section 4.5 and Appendix B.1). This establishes an inference benefit within each evaluated model without attributing it to another training run.

Table 6 and Table B.5 provide the separate training controls:

Without feature promptWeighted AUC@5 / 10 / 20RANSAC AUC@5 / 10 / 20
Supervision only37.98 / 59.33 / 75.4035.79 / 56.10 / 72.83
Detached gap gradients38.90 / 59.66 / 75.6136.34 / 56.56 / 73.26
Joint training38.90 / 59.98 / 75.6335.61 / 55.98 / 72.82

Supervision only removes all gap penalties during both training and inference and omits the prompt. Pinball therefore acts through shared correspondence features. This is a different model from the first Off row above, which was trained with gap weighting.

The independently trained controls are not a matched-budget decomposition. The same-model intervention isolates the contribution of gap weighting at inference.

Quantile supervision shapes shared features, and the learned gaps also carry useful information for weighted fitting. Table 3 still reports their combined addition, with the smaller prompt increment interpreted separately.

Revision. Sections 4.4-4.5 and Tables 5-6 (pp. 9-10) distinguish these comparisons. Appendix B.1/Table B.2 (p. 15) documents the same-model intervention; Appendix B.4/Table B.5 (pp. 16-17) defines the training controls and reports RANSAC results.

AuthorsOpenReview ↗

K3. Training variability and the size of the reported gains

Thank you for distinguishing repeatability across training runs from uncertainty over test pairs.

Concern. Do the smaller gains and the incremental feature-prompt benefit persist across independent seeds, with means and variability?

Response. Our paired native/RQR comparisons build on the corresponding public implementations, using the same correspondence inputs and evaluation settings within each pair. Each configuration is represented by one training run. The reported intervals quantify test-pair variability conditional on those trained models.

Table B.1 gives these representative paired AUC-point changes with 95% image-pair intervals:

RQR minus nativeWeighted AUC@5 gain [95% CI]Weighted AUC@10 gain [95% CI]Weighted AUC@20 gain [95% CI]
PointCN+0.30 [-0.26, +0.86]+1.19 [+0.50, +1.87]+1.25 [+0.64, +1.88]
BCLNet+3.77 [+2.98, +4.56]+2.79 [+2.11, +3.47]+1.39 [+0.89, +1.88]
GeoMoE+0.67 [-0.07, +1.40]+0.29 [-0.32, +0.89]+0.05 [-0.38, +0.48]

BCLNet's weighted gain has positive intervals at every tolerance. PointCN has positive intervals at 10 and 20 degrees, while its 5-degree interval includes zero. OANet, CLNet, DeMatch++ and GeoMoE have weighted intervals that include zero at all three thresholds. These are different levels of support within the evaluated test set.

The original RQR-F ablation remains in Table 3:

RQR-F configurationWeighted AUC@5Weighted AUC@10
Base37.6359.05
Quantile supervision and weighting38.9059.98
With feature prompt39.1560.12

Appendix B.1 states that the additional prompt gains have paired intervals that include zero. We therefore describe the prompt result as an observed increment, not an independently established repeatability gain. The same-model weighting intervention in Table 5 addresses a separate question: whether this trained model benefits from the gap penalties at inference.

Our study focuses on residual-supervision objectives and inference-time weighting across architectures and estimators. We will investigate repeatability across independent training seeds in future work.

Revision. Section 4.1 (p. 6), Section 4.4 (p. 9), and Appendix B/Table B.1 (pp. 14-15) identify what the intervals measure. The prompt interpretation is stated in Section 4.4 and Appendix B.1.

Reviewer_FsTZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Residual Quantile Regression (RQR) for two-view correspondence pruning, aiming to capture residual information beyond binary inlier confidence. RQR predicts ordered geometric residual quantiles and uses them either as training-time guidance for existing pruning networks or, in RQR-F, to refine features and fitting weights during inference. The method is evaluated across multiple correspondence pruning architectures, datasets, feature types, and geometric tasks, with additional robustness, transfer, ablation, and calibration analyses.

Critical strengths

1.The paper presents a clear and meaningful motivation by going beyond binary inlier supervision and explicitly modeling geometric residual information through residual quantiles. This provides a useful new perspective on correspondence pruning and gives the method a clear conceptual contribution beyond simply modifying the network architecture.
2. The experimental validation is relatively comprehensive, covering multiple pruning architectures, datasets, feature types, geometric tasks, robustness settings, and transfer scenarios. This broad evaluation strengthens the empirical support for the proposed residual-aware learning framework.

Critical weaknesses
  1. The related-work coverage is not sufficiently complete, particularly regarding recent studies on two-view correspondence pruning and geometric correspondence learning. This weakens the positioning of RQR relative to the latest developments in the field and makes it more difficult to assess the novelty of the proposed formulation.
  2. The empirical evidence does not yet fully establish the specific advantage and general effectiveness of quantile modeling. Under the matched all-interface setting, Huber and Laplace residual supervision outperform the three-quantile RQR variant, while corresponding final-interface comparisons are missing, suggesting that part of the gain may come from continuous residual supervision rather than quantile estimation itself. Moreover, the improvements are not consistently robust across hosts and evaluation settings: several gains are small, some settings show degradation, and the bootstrap intervals for multiple weighted comparisons include zero.
Questions for authors

1.Could the authors provide a matched comparison of final-interface RQR against final-interface Huber and Laplace supervision? This would help clarify whether the improvement is specifically due to quantile modeling or mainly to continuous residual supervision and supervision placement. 2.Could the authors clarify how RQR differs from and advances over recent work on two-view correspondence pruning and geometric reliability/residual modeling that is not sufficiently covered in the current related-work section? A clearer positioning would help assess the methodological novelty of the paper.

3 comments
AuthorsOpenReview ↗

Fs1. Residual targets under matched final-stage supervision

Thank you for identifying the need to separate residual-target choice from supervision placement.

Concern. Does final-stage RQR outperform matched final-stage Huber and Laplace, or are the gains explained by continuous residual supervision and where it is applied?

Response. The added final-stage controls show a quantile-objective advantage at the tightest weighted-pose tolerance under matched placement and capacity. The final-stage comparison uses the same BCLNet architecture, supervision location, three-output head capacity, initialization, classification and geometric objectives, and residual-loss coefficient (0.25). The objectives share the same training schedule and validation-selection budget. All auxiliary heads are removed at inference.

Tables 4 and B.4 give pose AUC (%) on YFCC-SIFT:

Final-stage objectiveWeighted AUC@5 / 10 / 20RANSAC AUC@5 / 10 / 20
Huber35.78 / 56.75 / 73.3734.82 / 54.99 / 71.88
Laplace36.19 / 57.24 / 73.8134.91 / 55.14 / 72.26
RQR, three quantiles37.13 / 57.80 / 73.8135.19 / 55.14 / 71.99

RQR's weighted AUC@5 gain is +1.35 [0.60, 2.12] over Huber and +0.93 [0.14, 1.72] over Laplace, with paired 95% image-pair intervals (Appendix B.3). These controls hold the prediction location fixed, so the difference cannot be attributed to moving supervision from all stages to the final stage within this comparison. The inference architecture is also unchanged because all auxiliary heads are removed.

The all-stage results in Table 4 remain: Huber and Laplace outperform three-quantile RQR there. At the final stage, RQR improves tight-angle weighted pose accuracy; at 20 degrees it matches Laplace to the reported precision. The corresponding RANSAC difference intervals include zero. The evidence therefore ties the quantile advantage to supervision placement and the geometric estimator.

The statistical fit of the outputs is a separate question. Table C.4 shows that Laplace has closer-to-nominal central coverage (77.58% versus 72.55%) on common retained matches. Consequently, our conclusion concerns the usefulness of quantile targets for learning correspondence features at the final stage, rather than universal superiority in residual-distribution estimation.

Revision. Section 4.4/Table 4 (p. 9) separates the two placements. Appendix B.3/Table B.4 (p. 16) reports the controls and intervals, while Appendix C.3/Table C.4 (p. 19) compares calibration. The conclusion (p. 10) reflects the tight-angle, weighted-estimation scope.

AuthorsOpenReview ↗

Fs2. Positioning RQR relative to recent geometric correspondence methods

Thank you for pointing out that the original related-work coverage made the methodological distinction difficult to assess.

Concern. How does RQR differ from recent correspondence-pruning and geometric reliability or residual-modeling methods?

Response. Appendix B.2 now compares RQR with NACNet (Moran et al., 2024), SC-Net (Lin et al., 2026), LeCoT (Dai et al., 2026) and BAG-Net (Wang et al., 2026). NACNet predicts displacement noise to correct match coordinates. SC-Net models spatial and cross-channel context, LeCoT changes the correspondence-pruning architecture, and BAG-Net aggregates forward consistency and reverse graph dependencies.

RQR adds supervision of ordered residual levels to correspondence learning. A scalar threshold classifier constrains the probability of falling below one residual threshold; pinball losses at several levels supervise residual location and spacing. Training-only integrations pass this supervision to the existing scores or features. RQR-F also uses the descriptor in a gated feature update and explicit gap-based weighting while preserving the observed coordinates. The contribution is this residual-target formulation and its integration into pruning and fitting, rather than a claim that prior methods ignore geometric reliability.

Table B.3 supplies the available quantitative comparison with recent complete systems, using the same 4,000 YFCC-SIFT pairs and pose-recovery procedure. AUCs are percentages:

MethodWeighted AUC@5 / 10 / 20RANSAC AUC@5 / 10 / 20
NACNet36.28 / 56.60 / 72.7933.13 / 52.90 / 69.79
LeCoT34.46 / 55.45 / 72.1735.29 / 55.50 / 72.37
RQR-F39.15 / 60.12 / 75.9436.08 / 56.73 / 73.38

NACNet's coordinate correction is preserved.

RQR-F is stronger in this evaluation under both estimators. This system-level comparison complements the within-network controls: Table 4 compares residual objectives at the same BCLNet stage, and Table 5 changes only gap weighting in the same trained RQR-F model. These controls connect the empirical argument to residual supervision and its use in geometric fitting, rather than to the name or age of the comparison architecture. SC-Net and BAG-Net are discussed methodologically; Table B.3 does not report results for them.

Revision. Appendix B.2 (pp. 15-16) discusses all four recent methods and Table B.3 (p. 16) gives the quantitative comparison. Sections 3.1-3.4 and Figure 2 (pp. 3-5) explain the distinct training and inference integrations; Tables 4-5 (p. 9) provide the targeted controls.

AuthorsOpenReview ↗

Fs3. Where the gains hold and where they are mixed

Thank you for pointing out that general effectiveness must be judged across architectures, estimators and transfer settings.

Concern. Several gains are small or uncertain, and some settings degrade. What scope of effectiveness does the evidence support?

Response. The results support architecture- and estimator-dependent benefits. We retain the unfavorable cases and distinguish observed improvements from gains with positive paired intervals.

Table B.1 gives representative weighted AUC-point differences and 95% image-pair intervals:

RQR minus nativeWeighted AUC@5 gain [95% CI]Weighted AUC@10 gain [95% CI]Weighted AUC@20 gain [95% CI]
PointCN+0.30 [-0.26, +0.86]+1.19 [+0.50, +1.87]+1.25 [+0.64, +1.88]
BCLNet+3.77 [+2.98, +4.56]+2.79 [+2.11, +3.47]+1.39 [+0.89, +1.88]
GeoMoE+0.67 [-0.07, +1.40]+0.29 [-0.32, +0.89]+0.05 [-0.38, +0.48]

BCLNet provides the strongest weighted evidence across all three tolerances. GeoMoE's positive weighted point estimates have intervals spanning zero; CLNet's weighted changes are mixed. OANet and DeMatch++ also have weighted intervals including zero. The intervals describe test-pair uncertainty for one trained model per configuration, not repeatability over training seeds.

Transfer also varies by estimator. In Table 2, OANet's weighted homography AUC@3/5/10px rises from 19.81/27.92/38.19 to 22.82/31.24/41.71, with RANSAC improvements as well. GeoMoE improves weighted DLT but its RANSAC AUC falls from 52.86/64.71/77.05 to 52.64/64.08/76.21. Figure 4 retains the cross-feature and cross-domain changes; Section 4.3 also reports the PointCN and OANet losses on YFCC-LoFTR.

The controlled results answer a more specific question than universal effectiveness: matched final-stage BCLNet supervision favors RQR over Huber and Laplace at weighted AUC@5 (Table 4), while removing gap weighting from the same trained RQR-F model reduces weighted accuracy (Table 5). These results support learning residual levels and using the predicted spacing in fitting under the evaluated conditions. They do not turn the mixed transfer or smaller backbone gains into uniformly positive findings.

Revision. Sections 4.2-4.3 and Tables 1-2/Figures 3-4 (pp. 7-8) retain architecture, estimator and transfer differences. Appendix B/Table B.1 (pp. 14-15) gives all paired intervals, and Sections 4.4-4.5 (pp. 9-10) distinguish the targeted control evidence.

Reviewer_FFyKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Residual Quantile Regression (RQR), an auxiliary objective for learned two view correspondence pruners. Besides the usual binary inlier label, a small head predicts several conditional quantiles of a log-transformed residual on the retained correspondences, trained with the pinball loss. Three integrations are proposed, (1) score guidance for PointCN/OANet/DeMatch++/GeoMoE, where a scalar "risk" built from the predicted median and 0.9-quantile is distilled into native scores via ranking/Smooth L1 losses, (2) shared feature supervision for CLNet/BCLNet, where the pinball loss back-propagates into the host's features, (3) RQR-F, a Transformer-style pruner in which the 3 quantile descriptor conditions features through a gated additive prompt and its gaps penalize weights in a differentiable weighted eight point solver. Paired experiments on YFCC show weighted pose AUC gains for most of the hosts, and transfer to SUN3D/SuperPoint/LoFTR and HPatches homography is mixed. Moreover, ablation on the RQR-F host gives better performance on AUC@5.

Critical strengths
  • The motivation is clean and correctly scoped.
  • The reporting is honest with paired bootstrap intervals, coverage diagnostics, and negative results are all included.
  • Training-only integrations add zero inference cost.
  • The paper uses a variety of hosts and settings for experiments.
Critical weaknesses
  • The paper’s main claim is that predicting several quantiles of the residual helps the pruner more than a binary inlier label does. To support that claim, the paper needs to show that quantiles do better than simpler ways of supervising the residual, such as predicting a single residual value. Table 4 runs this comparison on BCLNet, with supervision applied at all interfaces. In that setting, the simpler targets do better than RQR. Huber regression beats 3-quantile RQR, and Laplace regression also beats it at every threshold. However, the main BCLNet result uses supervision at the final interface only, and Huber and Laplace were not tested there. So the comparison that would directly support the paper’s claim is missing.
  • It is appreciated that the paper reports paired bootstrap confidence intervals. Those intervals show that many of the gains are not clearly different from zero. For OANet, CLNet, DeMatch++ and GeoMoE, the weighted-pose gains include zero at all three AUC thresholds. For PointCN, the gain at 5 deg. includes zero. With RANSAC, the gains for BCLNet and GeoMoE include zero. These intervals also capture only one source of variation, that is, which test pairs are sampled, for a single trained model. They do not capture how much results change when the same model is retrained with a different random seed.
  • The appendix states that geometry gradients reach the quantile head through the direct gap penalties. Since Eq 9 makes relative weight decrease monotonically with the gaps, the geometric loss directly pushes the gaps of useful matches down regardless of their actual residual spread. That is a plausible cause of the under-coverage in Appendix C. Then the descriptor is a jointly-learned weighting signal that is initialised or regularised by pinball loss, not an estimate of residual quantiles, and the paper's framing ("residual spread informs weights") should be qualified or tested with a stop-gradient variant.
Questions for authors
  • How does RQR-F compare with NACNet and with SC-Net/LeCoT/BAG-Net under your protocol? Does RQR still help when applied to the strongest current pruner?
  • Why is the predicted median absent from the explicit weighting in Eq 7?
  • How is the "predicted epipolar residual" defined on HPatches pairs, where the essential matrix is degenerate?
6 comments
AuthorsOpenReview ↗

F1. Final-stage controls distinguish the residual objectives

Thank you for identifying the missing comparison at the final prediction stage.

Concern. The all-stage results favor simpler residual targets; does RQR outperform Huber and Laplace at the stage used for the main BCLNet result?

Response. The matched final-stage comparison favors RQR at the tightest weighted-pose tolerance. The final-stage comparison uses the same BCLNet architecture, supervision location, three-output head capacity, initialization, classification and geometric objectives, and residual-loss coefficient (0.25). The objectives share the same training schedule and validation-selection budget. All auxiliary heads are removed at inference.

Tables 4 and B.4 report AUC (%) on YFCC-SIFT:

Final-stage objectiveWeighted AUC@5 / 10 / 20RANSAC AUC@5 / 10 / 20
Huber35.78 / 56.75 / 73.3734.82 / 54.99 / 71.88
Laplace36.19 / 57.24 / 73.8134.91 / 55.14 / 72.26
RQR, three quantiles37.13 / 57.80 / 73.8135.19 / 55.14 / 71.99

The weighted AUC@5 differences are +1.35 [0.60, 2.12] against Huber and +0.93 [0.14, 1.72] against Laplace (paired 95% image-pair intervals, Appendix B.3). Because the residual heads are removed at inference, the comparison tests how the objectives shape the correspondence features used by the native estimator.

The all-stage results in Table 4 remain: Huber and Laplace outperform three-quantile RQR there. At the final stage, RQR improves tight-angle weighted pose accuracy; at 20 degrees it matches Laplace to the reported precision. The corresponding RANSAC difference intervals include zero. The evidence therefore ties the quantile advantage to supervision placement and the geometric estimator.

We also separated task accuracy from distributional fit. On common retained matches, Table C.4 reports:

BCLNet objectiveCentral-80% coverageMean pinball loss
Laplace77.58%0.2143
RQR72.55%0.2151

Laplace has closer-to-nominal coverage and a slightly lower mean pinball loss; the interval for the mean pinball difference includes zero. RQR's final-stage pose advantage therefore supports its value as a learning objective for correspondence features, without implying better calibration than Laplace.

Revision. Section 4.4/Table 4 (p. 9) adds the final-stage controls and placement-dependent interpretation. Appendix B.3/Table B.4 (p. 16) provides both estimators and intervals; Appendix C.3/Table C.4 (p. 19) compares residual fit on common correspondences.

AuthorsOpenReview ↗

F2. Interpreting small gains and paired confidence intervals

Thank you for identifying both the intervals that include zero and the distinction between test-pair and training variation.

Concern. How should modest improvements be interpreted when paired intervals include zero and only one training run represents each configuration?

Response. We interpret each gain at its measured uncertainty level. Table B.1 reports all six networks and both estimators; the following weighted results illustrate the range of evidence (AUC-point changes and 95% intervals):

RQR minus nativeWeighted AUC@5 gain [95% CI]Weighted AUC@10 gain [95% CI]Weighted AUC@20 gain [95% CI]
PointCN+0.30 [-0.26, +0.86]+1.19 [+0.50, +1.87]+1.25 [+0.64, +1.88]
BCLNet+3.77 [+2.98, +4.56]+2.79 [+2.11, +3.47]+1.39 [+0.89, +1.88]
GeoMoE+0.67 [-0.07, +1.40]+0.29 [-0.32, +0.89]+0.05 [-0.38, +0.48]

The BCLNet weighted gains have positive intervals throughout. The smaller weighted changes for OANet, CLNet, DeMatch++ and GeoMoE have intervals that include zero, as does PointCN at 5 degrees. BCLNet and GeoMoE's RANSAC gain intervals include zero at all three tolerances. The paper retains these mixed outcomes.

The estimator difference is meaningful for BCLNet: weighted fitting retains continuous network weights, whereas RANSAC refits geometry from a filtered set. For CLNet/BCLNet, that set still depends on residuals under the network-predicted matrix. Appendix B explains this dependence, rather than treating RANSAC as entirely independent of the network's weighted estimate.

The paired native/RQR comparisons build on the corresponding public implementations and use the same correspondence inputs and evaluation settings within each pair. The bootstrap resamples image pairs with each trained model held fixed; it quantifies test-pair uncertainty for the reported single training run per configuration. Appendix B.1 also reports that the incremental prompt intervals include zero. We will investigate repeatability across independent training seeds in future work.

Our strongest residual-objective comparison is the matched final BCLNet stage, where the weighted AUC@5 advantage over both Huber and Laplace has positive paired intervals (Table B.4 and Appendix B.3). That conclusion is narrower than a uniformly reliable gain over all networks and estimators.

Revision. Section 4.1 (p. 6), Sections 4.2 and 4.4 (pp. 7 and 9), and Appendix B/Table B.1 (pp. 14-15) state the uncertainty target and retain the mixed solver-dependent results.

AuthorsOpenReview ↗

F3. Geometry gradients, detached gaps and residual calibration

Thank you for identifying the direct geometric-gradient path and its implications for interpreting the predicted gaps.

Concern. Could geometry shrink useful matches' gaps independently of residual spread, causing undercoverage? Does stopping that gradient change the result?

Response. The direct path does change the gaps. We now describe RQR-F's outputs as quantile-supervised residual descriptors and assess their fitting utility separately from their empirical coverage. Pinball acts throughout training; the geometric objective also reaches the descriptor through gap penalties and feature conditioning.

The detached-gap control keeps forward gap weighting, stops gradients through the quantile outputs entering the weighted solver, and omits the feature prompt. Geometry can still update shared features through the scoring branch. Tables 6 and B.5 report AUC (%):

Without feature promptWeighted AUC@5 / 10 / 20RANSAC AUC@5 / 10 / 20
Supervision only37.98 / 59.33 / 75.4035.79 / 56.10 / 72.83
Detached gap gradients38.90 / 59.66 / 75.6136.34 / 56.56 / 73.26
Joint training38.90 / 59.98 / 75.6335.61 / 55.98 / 72.82

The detached model retains weighted accuracy close to the joint reference at 5 and 20 degrees, with lower AUC@10; its RANSAC AUC is higher at all three tolerances. These observations show that useful pose estimation is possible with the direct gap path blocked. The independently trained variants have different training durations, so their AUC differences do not isolate the effect of detaching gradients.

Table C.3 measures direct-path gradients with other inputs held fixed. For the upper gap, the signed gradient ratio, ∑igi/∑i∣gi∣\sum_i g_i/\sum_i |g_i|, is +0.138 for geometric inliers in both prompt configurations and -0.104/-0.108 for outliers. These signs indicate aggregate pressure toward smaller inlier gaps and larger outlier gaps along the direct geometric-gradient path.

Residual diagnostics in Table C.2 provide a complementary test:

Feature promptAdjusted width/error correlationPinball increase after symmetrizingCentral-80% coverage
No0.3060.0012469.81%
Yes0.2970.0063875.74%

Width remains associated with median prediction error after linear rank adjustment for the predicted median and solver logits. Symmetrizing endpoints while fixing the median and total width increases pinball loss. These observations support residual-related information in the spacing; they are not a causal attribution of pose gains to asymmetry.

Central-80% coverage is 72.89% for detached gaps and 73.91% for supervision only (Section 4.5). Both under-cover. Thus removing the direct path alone does not restore nominal coverage. Pose accuracy and calibration remain distinct measured properties of the descriptor.

Revision. Section 3.4 (p. 5) distinguishes geometric paths; Section 4.5/Table 6 (p. 10) adds the control and coverage interpretation. Appendix B.4/Table B.5 (pp. 16-17) gives settings and RANSAC; Tables C.2-C.3 (p. 18) report residual information and direct-gradient diagnostics.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.