acceptodds
Under review as a conference paper at ICLR 2027

Geometry at the Last Mile: Beyond Local Appearance-Only Correspondence Search in Feature Matching

Abstract

Semi-dense matching has seen substantial progress, yet fine-stage localization remains predominantly driven by local appearance cues and lacks explicit pairwise geometry, leaving it vulnerable to ambiguous local evidence, especially under weak texture, repetitive patterns, and large viewpoint changes. We present GeMiL, an efficient local feature matcher that introduces shared image-pair geometry into the final stage of correspondence localization, moving beyond independent refinement driven solely by local appearance. At its core, a Structured Geometry Head, trained under direct supervision from camera pose, consolidates confident cross-view correspondences into a shared geometry prior. The resulting fundamental matrix acts as a gated and bounded geometric bias over fine candidates, suppressing inconsistent locations while reinforcing locally plausible ones. Since the quality of this geometry prior depends on the cross-view relation evidence from which it is inferred, we regularize the final cross-attention responses toward token-level correspondence targets through Correspondence-Grounded Cross-Attention. Sharp Relation Injection then propagates these correspondence-aware relation features across scales through content adaptive reassembly. Across diverse geometric tasks, GeMiL achieves consistently strong accuracy and leading efficiency among semi-dense matchers, running faster than ELoFTR with only 42% of its peak memory and 8.8% of its FLOPs.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Semi-dense matching, feature matching, correspondence learning, multi-view 3D reconstruction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_EcZKOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes GeMiL, a semi-dense feature matcher that injects pairwise geometry into the final correspondence-localization stage. A Structured Geometry Head predicts camera geometry from confident cross-view correspondences and converts the resulting fundamental matrix into a gated, bounded epipolar bias. Correspondence-Grounded Cross-Attention and Sharp Relation Injection improve the relation features used by the geometry head. The method is evaluated on relative pose, homography estimation, visual localization, and ETH3D reconstruction, with detailed efficiency measurements and ablations.

Critical strengths
  1. The central idea is simple and well motivated: local appearance remains ambiguous at the fine matching stage, while pairwise geometry can restrict plausible candidates. The gated and bounded bias is a sensible design because it uses geometry without allowing an inaccurate prediction to override appearance evidence completely.

  2. The evidence is strong across several downstream tasks. GeMiL leads the compared semi-dense methods on ScanNet pose estimation, HPatches, and ETH3D reconstruction while using 77 ms, 3.59 GiB, and 0.31 T FLOPs per MegaDepth pair. The component ablations and integrations into ELoFTR, EDM, and SLiM support both the individual design choices and the broader usefulness of the geometry head.

Critical weaknesses
  1. The gains over the strongest semi-dense baselines are limited on MegaDepth, and visual localization results are mixed, with a notable drop relative to ELoFTR on InLoc DUC2. Dense matchers remain more accurate in relative pose estimation. The paper would benefit from a clearer discussion of when the efficiency advantage justifies the remaining accuracy gap.

  2. The geometry head uses direct camera-pose supervision during training. This is reasonable for MegaDepth, but the dependence on calibrated geometric supervision may limit transfer to domains where accurate poses and intrinsics are unavailable. The current experiments do not test robustness to noisy training geometry or training on substantially different scene distributions.

Questions for authors
  1. How does performance change when the training poses or intrinsics are perturbed? This would clarify whether the geometry head remains useful with the imperfect supervision common in large web datasets.

  2. Can the authors characterize the subset of coarse correspondences for which the geometry gate activates, and report fine-localization accuracy separately for correspondences with active and inactive gates?

  3. Does the Structured Geometry Head still improve a matcher trained without metric camera intrinsics, or when only fundamental-matrix supervision is available?

4 comments
AuthorsOpenReview ↗

Response to Reviewer EcZK (1/2): Accuracy-efficiency trade-off and InLoc localization

We thank the reviewer for recognizing the motivation of geometry-guided fine matching. We clarify the efficiency trade-off, supervision requirements, and gate behavior below.

1. Accuracy–Efficiency Trade-Off (W1)

Accuracy Gap to Dense Matchers. The accuracy gap to dense matchers is not specific to GeMiL. On MegaDepth, all semi-dense methods in the comparison, including EDM and SLiM (current state-of-the-art semi-dense matchers), achieve lower pose AUC than DKM and RoMa. Their substantially lower runtime and FLOPs illustrate a shared accuracy–efficiency trade-off across these semi-dense approaches.

MethodTypeMegaDepth AUC@5 / @10 / @20 (%) ↑Time (ms) ↓Peak memory (GiB) ↓FLOPs (T) ↓
EDMSemi-dense57.5 / 73.2 / 84.283.458.240.50
SLiMSemi-dense57.9 / 72.8 / 83.5157.635.141.86
GeMiLSemi-dense57.4 / 73.2 / 84.377.473.590.31
DKMDense60.4 / 74.9 / 85.1572.989.663.91
RoMaDense62.6 / 76.7 / 86.3755.446.7915.30

Leading Efficiency among Semi-Dense Matchers. GeMiL maintains comparable accuracy to EDM while reducing peak memory by 56.4% and runtime by 7.2%. It is approximately 10× faster than RoMa and has the lowest memory and FLOPs among these methods, making it particularly useful under constrained latency and memory budgets. Under the same measurement protocol, GeMiL also reduces matching time by 44.9% (140.49 → 77.47 ms) and peak GPU memory by 58.4% (8.64 → 3.59 GiB) relative to ELoFTR.

When to Prefer GeMiL. GeMiL is particularly attractive when memory and latency are constrained: it achieves pose AUC within 0.1 percentage points of EDM at all three MegaDepth thresholds while reducing peak memory by 56.4%. When maximum pose accuracy is the primary objective and computational resources are sufficient, the evaluated dense matchers remain preferable.

InLoc DUC2. Localization recall (%) is reported below, with gains in percentage points in parentheses.

Geometric bias(0.25 m, 2°) ↑(0.5 m, 5°) ↑(1.0 m, 10°) ↑
Disabled51.167.280.2
Enabled (full model)54.2 (+3.1)70.2 (+3.0)81.7 (+1.5)

The bias improves DUC2 recall at all three thresholds. The remaining gap to ELoFTR may be related to differences in the underlying matching architecture.

We will discuss the remaining accuracy gap and the corresponding efficiency benefits in greater detail in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer EcZK (2/2): Geometric supervision, deployment and gate behavior

2. Geometric Supervision and Deployment (W2, Q1 and Q3)

The pretrained matcher requires only image pairs at inference, without target-domain GT poses or intrinsics. To address training scenarios with imperfect or limited geometric supervision, we additionally train two variants: one with noisy camera-parameter supervision and one with only fundamental-matrix supervision for the geometry head. Both are evaluated on MegaDepth-1500.

Noisy Supervision. We apply Gaussian perturbations with standard deviations of 1° for rotation and 5% for relative focal length, affecting only the head's direct rotation and FoV targets. Correspondence and Sampson supervision, as well as validation and testing, retain clean targets.

F-Only Supervision. We replace the head's camera-parameter and head-level Sampson losses with sign- and scale-invariant F regression, retaining the geometry head and fine-matching bias. Other matching losses remain unchanged, including the refinement Sampson loss. F and correspondence targets are still constructed from calibrated GT data; “F-only” refers specifically to the head's supervision.

Variant / supervisionAUC@5 (%) ↑AUC@10 (%) ↑AUC@20 (%) ↑
Full model57.473.284.3
Noisy rotation and focal-length supervision57.072.984.0
F-only head supervision56.972.783.9
w/o Structured Geometry Head55.971.783.7

Benefit with F-Only Head Supervision. Compared with the w/o Structured Geometry Head baseline, F-only head supervision achieves higher AUC at all three thresholds, with gains of approximately 1.0 / 1.0 / 0.2 percentage points based on the reported values. This supports the benefit of the geometry head and fine-matching bias with F-only direct supervision of the head.

Summary. Both variants retain performance close to the full model. Performance is largely preserved under the tested perturbations of the head’s direct rotation and FoV targets, with correspondence and Sampson supervision kept unchanged. The F-only results further support flexibility in the direct supervision of the geometry head.


3. Gate Activation and Fine-Localization Accuracy (Q2)

Evaluation Setup. Shares and confidence are pooled over coarse samples. Median EPE is pooled over GT-valid samples within each group, while accuracy is averaged per pair within each group, without RANSAC. EPE is in original-image pixels. PCK/MMA@1 denotes PCK@1 on MegaDepth and MMA@1 on HPatches.

DatasetGateShare (%)Mean coarse confidenceFine Median EPE (px) ↓PCK/MMA@1 (%) ↑
MegaDepthActive24.710.91660.72764.18
MegaDepthInactive75.290.84750.75563.84
HPatchesActive45.170.97130.42365.35
HPatchesInactive54.830.89780.88162.24

Summary. On both datasets, gate-active matches exhibit higher coarse confidence, lower median EPE, and higher PCK/MMA@1. This association supports the gate's role in selectively applying geometric guidance to more reliable correspondences.

Thank you for encouraging a closer examination of geometric supervision and gate behavior. We would be glad to discuss any remaining concerns about these findings and the scope of our conclusions.

Thank you for the response and additional experiments. It generally solved my concern. I maintain my score.

One small follow-up. For the same gate-active matches, how much does PCK/MMA@1 change if the epipolar bias is turned off at inference

Thank you for the follow-up. For the same gate-active matches, disabling the epipolar bias at inference reduces MegaDepth PCK@1 from 64.10% to 63.68% and HPatches MMA@1 from 65.12% to 61.98%. Thank you again for your time and thoughtful feedback.

Reviewer_Ca8cOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes GeMiL, a semi-dense matcher that injects learned pairwise geometry into fine-level refinement. A Structured Geometry Head predicts relative pose and intrinsics from coarse matches, and the resulting fundamental matrix gives a gated epipolar bias for fine matching. CGCA and SRI further improve the cross-view features. Experiments on relative pose, homography, visual localization and 3D reconstruction show competitive accuracy at low computational cost.

Critical strengths

The method is well motivated: shared geometry is a reasonable way to resolve local ambiguity in fine matching, and the gated, bounded penalty limits the effect of unreliable predictions. GeMiL is competitive at low runtime and memory, the evaluation covers several tasks, and the plug-in study suggests the head can benefit other matchers.

Critical weaknesses

Matching gains are confounded with the learned pose prior. The head is trained with ground-truth pose and intrinsics, and its fundamental matrix biases the fine scores at inference. The RANSAC pose is therefore computed from matches already conditioned on a learned pose prior, and a better pose AUC does not imply more accurate correspondences, since geometrically consistent matches are not necessarily correct. The stop-gradient at the predicted F separates the training of the two branches, but at inference the predicted geometry still shapes the matches.

Missing direct matching accuracy. Other methods are compared only through downstream tasks. EPE and PCK@1 appear only in an internal ablation, with small gains (0.755 to 0.727 px, 68.07% to 69.48%) and no seed variance.

Efficiency attribution. The speed and memory gains seem to come from the underlying lightweight design (low-resolution interaction, sparse matching, efficient fine regression) rather than from the proposed geometry modules.

Questions for authors

Since GeMiL uses pose supervision and a learned pose prior that other matchers do not have, a raw correspondence comparison independent of RANSAC is necessary: for example EPE, PCK@1/3/5 px and epipolar precision on MegaDepth-1500, and mean matching accuracy on HPatches, against ELoFTR, EDM, SLiM and RoMa at the same resolution and number of matches.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Ca8c (1/2): Direct correspondence accuracy on MegaDepth and HPatches

We thank the reviewer for recognizing the motivation and applicability of GeMiL. We agree that pose AUC alone is insufficient to assess the quality of the output matches and provide direct evaluations below.

1. Learned Geometry and Direct Matching Accuracy

We agree that higher pose AUC does not necessarily imply more accurate correspondences, and that stop-gradient does not remove the influence of predicted geometry at inference. We therefore evaluate correspondence accuracy directly against GT point locations, independently of RANSAC.

Evaluation Setup. For the following two cross-method comparisons, all methods use a 1152-pixel longest side and up to 1,000 highest-confidence matches per pair. Correspondences are evaluated against GT geometry without RANSAC. Errors are in original-image pixels. On MegaDepth, EPE is the Euclidean distance between the predicted target coordinate and the GT reprojection obtained from source depth, camera intrinsics, and relative pose. PCK measures the fraction of correspondences whose endpoint error falls below each threshold. Neither predicted geometry nor RANSAC is used to define the reference correspondences. Mean EPE and accuracies are averaged per pair. Epipolar precision (Epi.) uses squared symmetric epipolar error <1e-4 in normalized camera coordinates.

MegaDepth-1500. EPE is in pixels; PCK and Epi. are percentages. In the following two tables, the best results are bold and the second-best results are underlined.

MethodMean EPEPCK@1PCK@3PCK@5Epi.
ELoFTR3.837861.5793.83‾\underline{93.83}97.14‾\underline{97.14}98.59
EDM3.984964.6793.8296.9498.18
SLiM5.465866.11‾\underline{66.11}93.0795.1496.05
RoMa3.7947‾\underline{3.7947}59.1690.4395.1399.01
GeMiL3.188666.3694.6997.5198.76‾\underline{98.76}

HPatches (108 Sequences, 540 Pairs). MMA (%) measures agreement with the GT homography at each pixel threshold.

MethodMMA@1MMA@3MMA@5MMA@10
ELoFTR65.0891.2695.1696.66
EDM67.14‾\underline{67.14}91.7895.2796.61
SLiM63.6986.5490.6192.96
RoMa68.2291.92‾\underline{91.92}96.4198.46
GeMiL66.7292.0895.65‾\underline{95.65}96.86‾\underline{96.86}

Summary. GeMiL obtains the lowest mean EPE and the highest PCK@1/3/5 on MegaDepth. Among the evaluated semi-dense methods, GeMiL achieves the highest MMA at 3, 5, and 10 pixels on HPatches, while EDM performs better at the strict 1-pixel threshold. The plug-in comparisons below further show that adding the geometry head and bias consistently reduces EPE and increases PCK@3 across ELoFTR, EDM, and SLiM, supporting the effectiveness of geometric guidance for correspondence localization.

AuthorsOpenReview ↗

Response to Reviewer Ca8c (2/2): Training stability, efficiency attribution and plug-in results

2. Stability Across Training Seeds

Evaluation Setup. We agree that the modest gains in the original ablation could partly reflect training randomness. To address the concern about seed variance, we repeated training for both the full GeMiL model and the variant without the Structured Geometry Head, obtaining a total of three models trained with different random seeds for each variant. All metrics in this section are evaluated using all native output matches. For each variant, we computed the mean and variance of each direct matching metric on MegaDepth across the three runs. The table reports mean ± standard deviation, with the best mean results in bold.

MethodMed. EPE (px) ↓PCK@1 (%) ↑Samp. Med. ↓Epi.@1 (%) ↑
GeMiL (full)0.724 ± 0.00169.49 ± 0.100.3072 ± 0.000392.08 ± 0.02
w/o Structured Geometry Head0.774 ± 0.01367.88 ± 0.720.3186 ± 0.005990.48 ± 0.32

Summary. Averaged across the three training runs, the full model improves PCK@1 and Epi.@1 by 1.61 and 1.60 percentage points, respectively, over the variant without the Structured Geometry Head, while reducing median EPE and median Sampson error. It also achieves smaller standard deviations across all four metrics, supporting the consistency of the matching improvements across training seeds.


3. Efficiency Attribution and Plug-In Results

Summary. GeMiL builds on a baseline that balances accuracy and efficiency, and incorporates additional optimizations for efficiency. The geometry head improves matching quality at a very low computational cost. When integrated into ELoFTR, EDM, and SLiM, head computation takes only 8.69–9.70 ms per pair, accounting for 5.49%–9.13% of total runtime (see below). With the geometry head and bias, all three matchers achieve lower mean EPE, higher PCK@3, and higher pose AUC. Below, we report the plug-in pose results from Table 5 alongside the new direct matching-quality and runtime measurements.

Plug-In Evaluation Setup. Mean EPE and PCK are evaluated on MegaDepth-1500 using all native output matches, without top-k selection. Performance entries show baseline → + geometry head and bias, with signed changes (after minus before) in parentheses: pixels for EPE and percentage points for PCK and AUC. Changes are calculated before rounding. Head share is Head/Total runtime.

MatcherMean EPE (px) ↓PCK@3 (%) ↑AUC@5 ↑AUC@10 ↑Total (ms)Head (ms)Head share
ELoFTR4.9450 → 4.3691 (-0.5759)90.14 → 93.61 (+3.47)54.9 → 55.4 (+0.5)71.2 → 71.5 (+0.3)151.468.785.80%
EDM6.5917 → 4.5200 (-2.0717)86.87 → 93.79 (+6.92)55.4 → 56.5 (+1.1)71.2 → 72.4 (+1.2)95.218.699.13%
SLiM8.4495 → 7.2599 (-1.1896)84.68 → 88.66 (+3.98)55.4 → 56.7 (+1.3)69.2 → 70.9 (+1.7)176.729.705.49%

We appreciate your careful assessment of correspondence accuracy, training variability, and efficiency attribution. Your feedback has helped us strengthen the evaluation, and we welcome further questions or suggestions.

Thank you for the detailed response and additional experiments. I find the direct correspondence accuracy results strong, and I appreciate the clarification of the efficiency attribution. These results address my main concerns, and I will raise my rating.

One follow-up question: on average, how many correspondences does GeMiL produce per image pair on MegaDepth-1500 and HPatches? This is only for reference and would not affect my rating.

Reviewer_2h1sOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GeMiL, a semi-dense feature matcher that incorporates pairwise geometry into fine-level correspondence localization. Correspondence-Grounded Cross-Attention (CGCA) supervises cross-view attention using token-level correspondence targets, while Sharp Relation Injection (SRI) propagates relation features across scales. A Structured Geometry Head aggregates selected coarse matches to estimate relative camera pose and intrinsics, yielding a fundamental matrix that provides a gated geometric bias for fine-level matching. The method is evaluated on relative pose estimation, homography estimation, visual localization, and multi-view 3D reconstruction. Results show competitive accuracy and improved computational efficiency among semi-dense matchers, with ablation studies examining the contributions of individual components.

Critical strengths

S1. The paper addresses a clear limitation of appearance-based fine matching by incorporating image-pair geometry into final correspondence localization. The gated and bounded geometric bias provides a reasonable way to combine geometric constraints with appearance evidence.

S2. The experiments cover multiple geometric tasks, with ablations supporting the proposed components and cross-architecture evaluations demonstrating the applicability of the Geometry Head. GeMiL also achieves competitive accuracy with favorable runtime and memory efficiency among semi-dense matchers.

Critical weaknesses

W1. Limited methodological novelty. The framework largely builds on existing techniques, including correspondence-supervised attention, content-adaptive feature reassembly, and EDM-style fine regression. While their integration is effective, the paper does not clearly establish a substantial methodological advance beyond combining these established ideas.

W2. Unclear necessity of the learned Geometry Head. Although Table 4 shows better geometry estimation than RANSAC, this does not establish its added value in the final matching task. The paper lacks a direct comparison using RANSAC-estimated geometry with the same fine refinement strategy, leaving the necessity of the learned head unclear.

Questions for authors

Q1. How does RANSAC-estimated geometry perform when used with the same geometry-biased fine refinement? Could the authors compare final matching accuracy, pose estimation, and runtime?

Q2. Can the authors provide quantitative results under challenging matching conditions, such as repetitive textures and large viewpoint changes? In particular, does disabling the geometric bias improve the weaker localization results on InLoc DUC2?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 2h1s (1/2): Methodological contribution and cross-architecture matching accuracy

We thank the reviewer for recognizing the motivation of geometry-guided fine matching, the gated and bounded bias, and the cross-architecture applicability of the Geometry Head. We clarify the methodological contribution and provide direct comparisons below.

1. Methodological Contribution (W1)

Addressing Geometric Inconsistency in Semi-Dense Fine Matching with a Shared Geometric Prior. Our key contribution addresses a common limitation of semi-dense fine matching: independent, appearance-based local refinement can produce correspondences that are inconsistent with shared image-pair geometry. We address this by injecting a learned, shared geometric prior into local fine refinement. The Structured Geometry Head aggregates coarse correspondences to predict relative pose and intrinsics. The resulting fundamental matrix provides a gated, bounded bias over fine-level candidate scores. The plug-in results below demonstrate improvements in correspondence quality across multiple semi-dense matching architectures.

Feature Designs for Accurate Geometry Estimation. We design attention supervision and content-adaptive reassembly for matching models to support our overall geometry-guided refinement strategy. The former promotes geometrically valid correspondences, while the latter preserves localized correspondence evidence across scales, together providing reliable evidence for subsequent geometry estimation.

Cross-Architecture Evidence for Addressing a Common Limitation. To test whether the proposed mechanism addresses this common limitation across semi-dense architectures, we integrate the Geometry Head and geometric bias into ELoFTR, EDM, and SLiM. Table 5 reports their pose results, while Table 6 examines the contributions of the Geometry Head and geometric bias within GeMiL. We further evaluate direct correspondence quality on MegaDepth-1500 and HPatches.

The entries below show baseline → + Geometry Head/bias, with signed changes (after minus before) in parentheses: pixels for EPE and percentage points for PCK/MMA. Changes are calculated before rounding. Mean EPE is in original-image pixels; PCK and MMA are percentages, evaluated against GT geometry without RANSAC.

MatcherMegaDepth mean EPE ↓MegaDepth PCK@3 ↑HPatches mean EPE ↓HPatches MMA@3 ↑
ELoFTR4.95 → 4.37 (-0.58)90.14 → 93.61 (+3.47)11.70 → 10.87 (-0.83)84.94 → 90.74 (+5.80)
EDM6.59 → 4.52 (-2.07)86.87 → 93.79 (+6.92)12.41 → 10.39 (-2.03)86.07 → 92.08 (+6.01)
SLiM8.45 → 7.26 (-1.19)84.68 → 88.66 (+3.98)19.40 → 16.24 (-3.16)82.49 → 83.18 (+0.70)

Summary. The Geometry Head/bias variants achieve lower mean EPE and higher PCK@3/MMA@3 for all three matchers on both datasets. These consistent gains support our central motivation: local fine refinement across different semi-dense architectures can benefit from shared image-pair geometry. Together with the pose results in Table 5, they provide evidence that the proposed mechanism addresses this limitation across the evaluated architectures.

AuthorsOpenReview ↗

Response to Reviewer 2h1s (2/2): RANSAC comparison, challenging conditions and InLoc localization

2. RANSAC Geometry with the Same Fine Refinement (W2 and Q1)

Comparing RANSAC under the Same Refinement Strategy. We agree that the direct geometry-estimation comparison in Table 4 should be complemented by final matching accuracy, pose estimation, and runtime. We therefore compare three variants on MegaDepth-1500: A, appearance-only refinement with the geometric bias disabled; B, RANSAC-estimated F from the same top-512 coarse correspondences used by the Geometry Head, followed by the same geometry-biased fine refinement; and C, the complete GeMiL model with learned geometry. Mean EPE, PCK@3, and epipolar precision are evaluated on all native output matches.

SettingMean EPE (px) ↓PCK@3 (%) ↑Epipolar precision (%) ↑Pose AUC@5 / @10 / @20 (%) ↑
A: No geometric bias5.3086.9295.8156.0 / 72.1 / 83.5
B: RANSAC-F + bias5.1287.6796.0857.0 / 72.6 / 83.6
C: Complete GeMiL5.0088.2796.1857.4 / 73.2 / 84.3

RANSAC-based guidance also improves matching over appearance-only refinement. The learned head provides further gains of 0.60 percentage points in PCK@3 and 0.4 / 0.6 / 0.7 percentage points in pose AUC. Together with the runtime results below, these findings support its added value in correspondence refinement and tail latency under the evaluated setting.

Higher Accuracy and Lower P90 Latency at a Small Average Cost. We report average and 90th-percentile matching time for the same three settings.

SettingAverage time (ms) ↓P90 time (ms) ↓
A: No geometric bias68.3572.56
B: RANSAC-F + bias75.31197.11
C: Complete GeMiL77.4784.23

Compared with RANSAC-F, GeMiL reduces P90 runtime from 197.11 to 84.23 ms, a 57.3% reduction. The Geometry Head's fixed feed-forward computation avoids the data-dependent iterative search that can cause RANSAC runtime spikes on difficult image pairs. The lower P90 is consistent with this more predictable computation, alongside the gains in matching and pose accuracy.


3. Challenging Matching Conditions (Q2)

Challenging Scenarios. We evaluate viewpoint changes using the HardMatch Viewpoint and HPatches viewpoint subsets, and ambiguity from visually similar structures using HardMatch Doppelgänger. All methods are evaluated using all native output matches, without top-k selection. GeMiL denotes the full model with the geometric bias enabled.

We report fundamental-matrix mAA@5/10 (%) on HardMatch and direct matching MMA@3/5 (%) on HPatches viewpoint. MMA uses GT homography transfer errors in original-image pixels.

HardMatch.

MethodViewpoint mAA@5 / @10 ↑Doppelgänger mAA@5 / @10 ↑
GeMiL8.66 / 12.9313.81 / 19.53
EDM7.63 / 11.6615.44 / 21.03
SLiM7.10 / 10.8411.33 / 16.55

HPatches Viewpoint.

MethodMMA@3 (%) ↑MMA@5 (%) ↑
GeMiL79.7186.89
EDM79.5386.68
SLiM74.7082.44

Summary. Using all native output matches, GeMiL surpasses EDM and SLiM in fundamental-matrix mAA@5/10 on HardMatch Viewpoint and direct matching MMA@3/5 on HPatches viewpoint, supporting its effectiveness under viewpoint changes. On Doppelgänger, GeMiL outperforms SLiM but trails EDM, indicating that ambiguity from visually similar structures remains challenging.


4. Effect of Geometric Bias on InLoc DUC2 (Q2)

Disabling the Geometric Bias Reduces DUC2 Localization Recall. Using the same weights and switching the geometric bias on or off only at inference, disabling the bias reduces localization recall at all three thresholds:

DUC2 settingRecall at (0.25 m, 2°)Recall at (0.5 m, 5°)Recall at (1.0 m, 10°)
GeMiL, bias off51.1%67.2%80.2%
GeMiL, bias on54.2%70.2%81.7%

Enabling the bias improves recall by 3.1 / 3.0 / 1.5 percentage points, supporting a positive contribution from geometric refinement on DUC2. ELoFTR achieves 59.5 / 82.4 / 87.0%, so a performance gap remains. However, removing the geometric bias further lowers GeMiL's recall and does not resolve that gap.

Thank you again for your thoughtful questions about the contribution of geometric refinement and its performance in challenging scenes. We would be happy to clarify any remaining points and continue the discussion.

AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer 2h1s,

As the discussion deadline approaches, we would like to follow up on our responses, which include additional experiments and clarifications addressing your questions. We would appreciate hearing whether our responses have addressed your concerns or if any points need further clarification, so that we can respond before the discussion closes.

Thank you for your consideration.

AuthorsOpenReview ↗

Hoping to Hear Your Thoughts Before the Discussion Closes

Dear Reviewer 2h1s,

We have provided additional experiments and clarifications addressing your concerns and incorporated all the key experiments you requested into the revised manuscript.

With less than a day remaining in the discussion period, we would be grateful if you could review our responses and take another look at our work in light of these revisions. If you have any further questions, we would be happy to discuss them and provide clarification.

Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers: Summary of Manuscript Revisions

We thank all reviewers for their constructive comments. The revised manuscript incorporates four additional experimental studies and related clarifications from our responses.

Reviewer 2h1s

  • Evaluations under large viewpoint changes and ambiguity from visually similar structures on HardMatch Viewpoint, HardMatch Doppelgänger, and HPatches Viewpoint, comparing GeMiL with EDM and SLiM (Sec. 4.5, Table 3).
  • Plug-in correspondence-accuracy evaluation of the Structured Geometry Head and geometric bias in ELoFTR, EDM, and SLiM on MegaDepth-1500 (App. E, Table 8).
  • Geometric-bias ablation on InLoc DUC2 using the same trained weights, with the bias enabled or disabled only at inference (App. F, Table 9).

Reviewer Ca8c

  • Direct correspondence-accuracy evaluation of the plug-in variants using mean EPE and PCK@3 on all native output matches, alongside pose AUC@5 (App. E, Table 8).
  • Runtime breakdown for the augmented ELoFTR, EDM, and SLiM models, separately reporting total matching time and Geometry Head computation time (App. E, Table 8).

Reviewer EcZK

  • Geometric-bias ablation on InLoc DUC2 to assess its contribution to indoor localization performance (App. F, Table 9).
  • Experiments with noisy rotation and focal-length supervision and F-only head supervision on MegaDepth-1500 (App. G, Table 10).
  • Clarification that inference requires only image pairs and that F-only supervision refers specifically to the head's direct supervision, with targets still derived from calibrated ground-truth data (App. G).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.