- 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
-
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.
-
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
-
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.
-
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
-
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.
-
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?
-
Does the Structured Geometry Head still improve a matcher trained without metric camera intrinsics, or when only fundamental-matrix supervision is available?
4 comments
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 / supervision | AUC@5 (%) ↑ | AUC@10 (%) ↑ | AUC@20 (%) ↑ |
|---|---|---|---|
| Full model | 57.4 | 73.2 | 84.3 |
| Noisy rotation and focal-length supervision | 57.0 | 72.9 | 84.0 |
| F-only head supervision | 56.9 | 72.7 | 83.9 |
| w/o Structured Geometry Head | 55.9 | 71.7 | 83.7 |
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.
| Dataset | Gate | Share (%) | Mean coarse confidence | Fine Median EPE (px) ↓ | PCK/MMA@1 (%) ↑ |
|---|---|---|---|---|---|
| MegaDepth | Active | 24.71 | 0.9166 | 0.727 | 64.18 |
| MegaDepth | Inactive | 75.29 | 0.8475 | 0.755 | 63.84 |
| HPatches | Active | 45.17 | 0.9713 | 0.423 | 65.35 |
| HPatches | Inactive | 54.83 | 0.8978 | 0.881 | 62.24 |
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.
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.
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.
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.