acceptodds
Under review as a conference paper at ICLR 2027

GACLoc: Ground-Anchor Geometry Modeling for Cross-View Localization

Abstract

Cross-view geo-localization estimates the planar camera pose of a ground image with respect to geo-referenced satellite imagery. Recent correspondence-based cross-view localization methods enable interpretable camera pose recovery through explicit ground–satellite matches. However, learning such correspondences remains challenging because training data usually provide only camera-pose supervision, without pixel-level ground–satellite correspondence annotations. Cross-view correspondence further exhibits an inherent geometric asymmetry: under a known ground sampling distance (GSD), equal-sized satellite tokens correspond to approximately fixed metric regions, whereas perspective ground tokens correspond to depth-dependent regions whose size, shape, and orientation vary substantially. Moreover, in repetitive urban scenes, visual similarity alone is insufficient to establish reliable correspondences, as incorrect satellite locations can receive high matching scores. To address these issues, we propose GACLoc: Ground-Anchor Geometry Modeling for Cross-View Localization (GACLoc), a geometry-aware correspondence framework with two key components: Anchor Geometry Representation (AGR) and Geometry-Aware Local Correspondence Refinement (GALCR). AGR models the metric region geometry of ground anchors and uses it to construct geometry-based candidates and geometric features, while GALCR fuses visual and geometric features to obtain more reliable soft correspondences for weighted analytic pose recovery. Experiments demonstrate competitive localization accuracy under challenging settings, including cross-area evaluation and unknown orientation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
cross-view localization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Fw5AOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds on Loc² for fine-grained cross-view localization. Stage I follows Loc² and adds a region loss that trains each ground token to match a small Gaussian region on the satellite image instead of a single point. Stage II uses monocular depth to turn each ground token into a region on the ground plane and places it on the satellite image using the coarse pose, then retrieves nearby candidates, and reranks them with a cross-attention module before a final weighted procrustes solve. The method shows the best mean errors on VIGOR and KITTI under unknown orientation but is weaker when orientation is known or heading noise is small.

Critical strengths

1- The refinement step is sensible and gives consistent gains under unknown orientation on both datasets which is the hardest and most practical setting.

2- The paper writing and overall presentations are clear.

Critical weaknesses

1- The main idea isn’t really supported by the paper’s own results. The paper argues that modeling the size, shape, and orientation of each token’s region is what improves matching. But in Table 8, replacing the region covariance with a simple circle gives the same final pose and even better candidate recall without any retraining. This means the gain comes mainly from searching near where the coarse pose predicts the match which is standard guided matching. The region loss, a large part of the total improvement, is also never compared with a simple isotropic target. Without this evidence the contribution looks like an extra refinement pass on top of Loc² than a new idea.

2- The paper doesn’t show that Stage II is worth its cost. Stage II reduces the mean error by about half a meter on VIGOR but it adds a full monocular depth model at test time and cross-attention over 1024 anchors with up to 32 candidates each. No runtime, memory, or FLOPs are reported either for Stage I alone or for the full model compared with Loc². The gains also show up mainly under unknown orientation. With known orientation or small heading noise, the method falls behind FG², Loc², or HC-Net, so I think this needs justifications. Also, part of the improvement may come from setup differences (a newer depth model, cubemap processing, RANSAC in the main results only), and this also needs to be checked to have a fair comparison with baselines.

Questions for authors

Clarifications:

1- What are the inference time, memory, and FLOPs of Stage I alone, the full model, and Loc² on the same hardware including the depth model? 2- In the known-orientation and ±10° settings, does the solver use the orientation prior, or is yaw estimated freely? This could explain the weaker results there. 3- How does your Stage I differ from Loc² (depth model, cubemap processing, RANSAC)? Your visual baseline doesn’t match Loc² reported numbers. 4- Since a simple circle works as well as the full covariance in Table 8, what evidence shows that region shape matters for reranking or for the region loss?

Suggestions:

1- An accuracy-runtime plot comparing Stage I, the full model, and prior methods would make the value of Stage II much clearer. 2- A simple guided-matching baseline such as rerunning Stage I matching within a window around the coarse-pose prediction would show what the learned reranker adds.

5 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer Fw5A

Dear Reviewer,

We hope everything is going well for you. We are truly grateful for your thoughtful review and constructive comments on our work. We sincerely hope that our responses have helped address your concerns and would appreciate hearing whether any points still require clarification. If you have any further questions or suggestions, we would be delighted to discuss them with you. Thank you again for your valuable feedback and for the time and effort you have devoted to evaluating our manuscript.

Area_Chair_yWcUOpenReview ↗

Invitation to Continue the Discussion

Dear Reviewer,

Thank you for your valuable feedback on this submission. Since we are now in the discussion phase, I would be grateful if you could engage with the authors’ responses and continue the conversation if there are outstanding issues that would benefit from clarification. Your participation will help ensure a constructive and thorough evaluation.

Many thanks again for your time and consideration.

Best regards,

Area Chair

Thanks for the reply and the new experiments. They address my main concerns. One question first: your new ablations show the simple circle works as well as or better than the ellipse. So the gain seems to come from the region size and the reranker, not the shape. Can you justify this here and could you update the paper’s claims to reflect this?

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

This paper addresses fine-grained cross-view localization, the task of estimating the planar 3 dof pose of a ground-level image by matching it to a local aerial image. It proposes GACLoc, a two-stage framework that first estimates a coarse pose through visual matching (following Loc2), then uses depth-derived region geometry to retrieve and refine satellite correspondence candidates. The camera pose is computed using weighted Procrustes alignment as in prior works. Experiments on VIGOR and KITTI show competitive localization accuracy, particularly under unknown orientation.

Critical strengths
  • Matching local features between ground and aerial images with only camera pose as supervision is a challenging but interesting task

  • Considering geometry for ground-aerial local feature matching is a reasonable design.

  • The performance under the unknown orientation setting is strong.

  • The paper provides informative diagnostic experiments on candidate coverage, correspondence inlier ratios, and sensitivity to the initial pose.

Critical weaknesses
  • The proposed method is largely based on Loc2, with Stage II refinement appearing to be the main contribution. I am unsure whether this constitutes a sufficiently substantial technical advance.

  • The authors identify the different BEV footprints of ground and aerial patches as a key challenge. However, Table 8 provides limited support for the importance of modeling the ground patch’s shape in candidate retrieval, as circular and elliptical regions yield nearly identical performance.

  • It is unclear how the refinement performs when the depth-derived geometry contains errors. Could the authors evaluate its robustness to noisy depth estimates?

  • The contribution of learned confidence weighting is not isolated. Seems Stage II changes both the correspondences and their weighting and selection.

  • How much computational overhead is introduced compared to Loc2?

  • Fig. 3 is too small to read (this did not impact my rating)

Questions for authors
  • Given the nearly identical results for circular and elliptical retrieval in Table 8, what evidence supports the importance of modeling the ground-region shape beyond its center?

  • Could the authors evaluate sensitivity to depth noise, including whether Stage II retains its advantage over Stage I as depth errors increase?

  • What inference-time and memory overhead does the method introduce compared with Loc² on the same hardware?

5 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer bAPP

Dear Reviewer,

We hope you are doing well. We sincerely thank you for your thorough review and valuable suggestions regarding our manuscript. We would greatly appreciate knowing whether our responses have sufficiently clarified the issues you raised. Should any questions or concerns remain, we would be pleased to provide additional explanations and continue the discussion with you. Once again, thank you for the time and effort you have dedicated to reviewing our work.

Area_Chair_yWcUOpenReview ↗

Invitation to Continue the Discussion

Dear Reviewer,

Many thanks for your careful review. I would like to kindly encourage you to review the authors’ rebuttal and, where appropriate, leave a follow-up comment. Further discussion of any remaining points would be valuable in clarifying the strengths and limitations of the work.

Thank you very much for your continued effort.

Best regards,

Area Chair

I thank the authors for their rebuttal, which answers my question about the performance difference between the isotropic and anisotropic settings. The authors present region modeling as their main contribution because it preserves information about the physical area represented by each ground-view patch. On the datasets tested, however, this component appears to yield only modest gains. Other reviewers seem to share this concern.

How did the authors choose the ±5% perturbation for the depth noise experiment? Is it representative of the relative errors that depth estimation models typically produce on these datasets?

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

The paper addresses 3-DoF ground-to-satellite localization. Extending depth-lifted local matching, it models each ground anchor as a bird’s-eye-view region with a centroid and covariance, retrieves satellite candidates by Mahalanobis distance, and reranks them with a cross-attention module fusing visual and geometric features before weighted Procrustes. It reports lower mean errors than its base method on VIGOR and KITTI under unknown orientation.

Critical strengths

The strengths of the paper are listed as follows:

  • Well written and easy to follow. Figure 2 mirrors Sections 3.1–3.3, each module ends by listing what it hands to the next (Sections 3.1–3.2), and the coarse-pose dependence is stated where the numbers appear. The main text stays self-contained at the level of inputs and outputs.
  • Well-motivated geometric asymmetry. Perspective tokens cover depth-dependent regions while satellite tokens cover fixed ground-sampling-distance cells. Summarizing an anchor by a projected covariance and gating candidates by Mahalanobis distance is a principled, cheap way to add this to point-based matchers such as Loc2.
  • Clear gains in the hardest settings. Mean errors fall from Loc2’s 4.23 m / 11.67 degrees to 3.32 m / 9.49 degrees on cross-area unknown-orientation VIGOR and from 1.85 / 1.31 m to 1.39 / 0.90 m on KITTI ±180 degrees same-area.
Critical weaknesses

The weaknesses of the paper are listed as follows:

  • Incremental novelty and missing related work. GALCR is a LoFTR-style fine stage (softmax-expectation over a local candidate set) [1] whose window is set by a covariance, much as ASpanFormer sizes its span by uncertainty [2]. Residual similarity refinement with a confidence gate and dustbin already appears in ViewBridge [3], which is uncited.
  • No efficiency or backbone-matched comparison. Stage II adds a Depth Anything 3 pass, cross-attention over 1,024 anchors × 32 candidates and coarse-solve RANSAC (Section 4.2), but no latency, memory or FLOPs are reported.
  • The core claim that region shape and orientation matter is unsupported. Only retrieval separates center from region: swapping in an isotropic covariance changes mean pose error by just 0.011 m and 0.06 degrees and raises center recall (70.3% vs. 66.2%; Table 8), and Appendix A.4 concedes “limited sensitivity”. The anisotropic terms of Eqs. 7–8 are never ablated, nor is the region loss (Table 3) compared with an isotropic target. On anchors above 10 square meters, where regions should matter most, refinement worsens the center error by 3.04 m on average.

[1] Sun, Jiaming, et al. "LoFTR: Detector-free local feature matching with transformers." arXiv preprint arXiv:2104.00680 (2021).

[2] Chen, Hongkai, et al. "Aspanformer: Detector-free image matching with adaptive span transformer." European conference on computer vision. Cham: Springer Nature Switzerland, 2022.

[3] Xia, Panwang, et al. "ViewBridge: Revisiting Cross-View Localization from Image Matching." arXiv preprint arXiv:2508.10716 (2025).

Questions for authors

It would be great if the authors could focus on my concerns in the Weaknesses section in the rebuttal.

3 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer hUEn

Dear Reviewer,

We hope this message finds you well. We sincerely appreciate your careful review and constructive feedback on our manuscript. We would be grateful to know whether our responses have adequately addressed your concerns. If any questions or uncertainties remain, we would be more than happy to provide further clarification and engage in additional discussion. Thank you once again for your valuable time and effort in evaluating our work.

Area_Chair_yWcUOpenReview ↗

Invitation to Continue the Discussion

Dear Reviewer,

Thank you again for your thoughtful review and the time you have devoted to this submission. As the discussion period is underway, I would greatly appreciate it if you could consider the authors’ response and engage further in the discussion should you have any remaining questions or concerns. A brief follow-up on whether the response addresses your points would also be very helpful.

Best regards,

Area Chair

AuthorsOpenReview ↗

Thanks to Reviewer hUEn

Dear reviewer hUEn, Thank you very much for your support, which means a great deal to us. We have incorporated your suggestions into the revised manuscript. We sincerely appreciate the time and effort you have devoted to reviewing our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.