acceptodds
Under review as a conference paper at ICLR 2027

Beyond Local Refinement: Wide-Area Precise Cross-View Localization

Abstract

Fine-grained cross-view localization aims to precisely localize a ground camera against geo-referenced aerial imagery, but current methods typically assume a narrow geographic prior, substantially limiting their applicability when upstream localization remains coarse. To bridge this gap, we study this task over much larger aerial regions. However, this seemingly simple extension fundamentally changes the localization problem: the growing disparity in spatial extent between ground and aerial views makes reliable cross-view matching difficult to establish directly, while the increasing number of look-alike locations in larger regions introduces substantial visual ambiguity. To address these challenges, we propose RingMatch, a search-and-verification framework for wide-area cross-view localization. RingMatch first leverages predictive uncertainty to uncover plausible locations and progressively zooms into their surroundings to obtain increasingly fine-grained evidence, yielding a compact set of promising candidates. We then introduce the *depth ring*, a novel geometric representation that reveals subtle variations in the spatial layout around a location and serves as a view-shared spatial signature for an LVLM verifier to resolve residual ambiguity. To benchmark this setting, we construct VIGOR-Pro by composing VIGOR aerial tiles into contiguous regions with spatial extents of up to . Experiments show that RingMatch maintains strong localization performance as the aerial region expands, consistently outperforming representative end-to-end and retrieval-based methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
fine-grained cross-view localization, spatial reasoning, coarse-to-fine search, cross-view geometry

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_YN3TOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates wide-area fine-grained cross-view localization, extending conventional local refinement to substantially larger aerial regions. It identifies two key challenges: sparse ground–aerial correspondences and visual ambiguity among look-alike locations. To address these challenges, the authors propose RingMatch, combining uncertainty-guided multi-hypothesis branch-and-zoom search with depth-ring-grounded LVLM verification to distinguish plausible candidates. The authors also introduce VIGOR-Pro, constructed by composing VIGOR aerial tiles into contiguous regions of up to 400 m. Experiments across same-area and cross-area settings demonstrate improved localization accuracy over the end-to-end and retrieval-based baselines.

Critical strengths
  1. A novel research perspective for fine-grained CVGL: The paper identifies a meaningful limitation of existing fine-grained cross-view localization methods: their reliance on narrow aerial priors. By extending localization to wide-area regions with coarse upstream priors, the paper studies a more realistic setting where direct local alignment becomes insufficient. The novel perspective of CVGL exposes challenges that are hidden under conventional small-region evaluation protocols.

  2. A well-designed search-and-verification framework: RingMatch presents a coherent framework aligned with the challenges. The multi-hypothesis branch-and-zoom search addresses uncertainty under sparse correspondence, while the depth ring representation provides explicit geometric evidence for resolving look-alike candidates. The separation between exploration and verification makes the framework interpretable and modular.

Critical weaknesses
  1. Unclear Contribution of Depth Ring: In Table 3, the contribution of the proposed depth ring representation itself is still partially entangled with the LVLM verifier and search initialization. The full verifier benefits from both additional geometric evidence and initialization from the trained search model. However, it remains unclear how much improvement comes from the depth-ring evidence. The authors should provide a more direct comparison with alternative geometric representations or non-LVLM geometric verification to better establish the unique contribution of depth rings.

  2. Limited Analysis on Computation: The extra modules like iterative search and LVLM-based verification are introduced in the framework. Since the task requires real-time performance, the authors should provide a sufficient analysis of the computational cost of this pipeline, which would strengthen the practical significance of the method.

Questions for authors

Q1. Contribution of Depth Rings

Table 3 demonstrates the benefit of depth-ring evidence, but the improvement is combined with the verifier and search initialization. Could the authors provide a comparison with alternative geometric representations or non-LVLM geometric matching approaches to better isolate the contribution of the depth ring itself?

Q2. Search and Verification Efficiency

RingMatch introduces multiple search stages and LVLM-based verification. Could the authors report the average inference cost, including the number of search model evaluations, depth-ring predictions, and LVLM verification calls, and discuss the accuracy–efficiency trade-off compared with existing localization pipelines?

Q3. Generalization Beyond Known Orientation

The current experiments focus on the known-orientation setting, where ground panoramas have a known heading relative to the aerial image. Could the authors discuss how RingMatch would extend to unknown-orientation scenarios, where depth-ring alignment requires additional yaw estimation?

4 comments
AuthorsOpenReview ↗

Response to Reviewer YN3T (Part 1/2)

We sincerely thank the reviewer for the positive assessment and constructive suggestions. We address the questions below with additional experiments and clarifications.

1. Contribution of depth rings (Q1)

As reported in Table 2 of our submission, our Ring-only baseline using Pearson correlation already achieves 69.42% verification accuracy at the 400 m scale without learned LVLM verification, demonstrating the discriminative value of depth-ring representations. To further isolate their contribution, we introduce two additional controls: a probabilistic NLL ring-matching baseline and a Search-initialized Verifier without depth-ring inputs. NLL matching selects the candidate with the lowest negative log-likelihood of ground-ring depths under its predicted aerial depth distributions.

Verification ablation under same-area evaluation at the 400 m scale.

MethodSearch init.Depth ringALL mean (m) ↓VFY mean (m) ↓VFY acc. ↑
Pearson ring matchingN/AYes15.2367.8869.42%
NLL ring matchingN/AYes13.8557.8775.56%
Verifier w/o ringsYesNo14.5362.7974.44%
Verifier w/ ringsYesYes12.3647.1282.81%

ALL denotes all test queries, while VFY denotes the subset requiring verification.

Depth rings provide discriminative geometric evidence independently of learned LVLM verification. Non-LVLM ring matching achieves up to 75.56% verification accuracy, demonstrating the utility of depth-ring evidence without learned verification. With Search initialization held fixed, adding depth rings further improves LVLM verification accuracy from 74.44% to 82.81% and reduces localization error, confirming their complementary contribution to the complete system.


2. Search and verification efficiency (Q2)

We report per-query Search evaluations, depth-ring computations, and Verifier calls, together with stage-wise and end-to-end latency. All experiments use same-area evaluation at the 400 m scale on a single NVIDIA A100 80 GB GPU, processing one query at a time after warm-up. Timing includes image preparation and all inference stages. We also vary the search depth (L=2, 3, and 4) to demonstrate the flexibility of our method in balancing localization accuracy and inference efficiency.

Mean RingMatch workload per query under same-area evaluation at the 400 m scale.

Search depthSearch nodesCoordinate samplesGround ringsAerial ringsVerifier callsVerified queries (%)
L=45.0480.670.1370.2740.13713.7
L=33.5757.050.1520.3040.15215.2
L=22.1934.980.1570.3140.15715.7

Search nodes correspond to Search model evaluations; coordinate samples count stochastic coordinate predictions. Values are averaged over all queries.

Mean stage latency per query under same-area evaluation at the 400 m scale.

MethodImage prep. (ms)Search (ms)Depth-ring prep. (ms)Verification (ms)
RingMatch (L=4)542.91,242.449.345.6
RingMatch (L=3)541.5994.352.150.6
RingMatch (L=2)542.0746.758.058.2

Image prep. includes ground and aerial image I/O and preprocessing. Depth-ring prep. includes ground depth estimation, aerial ring prediction, and rendering. Stage latencies are averaged over all queries.

End-to-end latency and localization error under same-area evaluation at the 400 m scale.

MethodLatency, mean / P95 (s)Localization error, mean / median (m)
Qwen3-VL-4B (End-to-End)0.964 / 1.01623.03 / 3.93
Sample4Geo + FG²2.149 / 2.70015.53 / 4.72
RingMatch (L=4)1.881 / 2.86012.36 / 1.88
RingMatch (L=3)1.639 / 2.46012.85 / 2.12
RingMatch (L=2)1.405 / 2.13914.25 / 2.83

For Sample4Geo + FG², timing includes image preparation, ground and aerial encoding, retrieval, and FG² refinement.

RingMatch achieves a favorable and flexible accuracy–efficiency trade-off. Across all tested search depths, RingMatch achieves both lower mean localization error and lower mean inference latency than Sample4Geo + FG². At L=4, it reduces mean error by 20.4% and mean latency by 12.5%, although its P95 latency is slightly higher. Despite its multi-stage design, additional verification overhead remains low: only 13.7% of queries require verification, with depth-ring preparation and verification adding just 94.9 ms per query on average. Furthermore, reducing the search depth provides additional latency savings with a modest increase in localization error, allowing flexible adjustment to different computational budgets.

AuthorsOpenReview ↗

Response to Reviewer YN3T (Part 2/2)

3. Generalization to unknown orientation (Q3)

RingMatch can be extended to unknown-orientation localization by jointly predicting location and heading. Specifically, the Search model can be trained to generate [x,y,θ][x,y,\theta] instead of [x,y][x,y]. During the branch-and-zoom search, heading predictions within each hypothesis cluster can be aggregated through circular averaging at each level, allowing both location and heading to be progressively refined. The estimated heading can then be used to align the ground and aerial depth rings during verification. We leave this extension as future work.

Depth-ring matching remains robust under heading uncertainty. Our previous yaw-free ring-matching experiments evaluate Pearson and NLL matching without known heading by searching over circular shifts to find the best alignment between ground and aerial rings. As shown in the table below, NLL matching achieves 73.55% verification accuracy, only 2.01 percentage points below the known-heading setting. This small performance drop indicates that depth-ring representations retain strong discriminative capability even when the relative orientation is unknown.

Verification accuracy with known and unknown heading under same-area evaluation at the 400 m scale.

Ring matchingKnown headingUnknown heading
Pearson69.42%66.01%
NLL75.56%73.55%

We hope these additional experiments and analyses satisfactorily address the reviewer's questions. We will incorporate these additional results into the revised manuscript to further strengthen our work. We thank the reviewer again for the valuable feedback.

I appreciate the authors’ detailed responses and additional experiments, which have addressed my remaining concerns regarding the contribution of depth-ring representations, efficiency, and robustness to unknown orientations. The new experimental results further strengthen the effectiveness and practicality of RingMatch. I believe this paper makes a valuable contribution to fine-grained CVGL, and I have no further concerns and continue to maintain my positive score.

Thank you for your encouraging feedback and positive assessment of our work. We are delighted that our additional experiments and clarifications have helped address your concerns. We sincerely appreciate your recognition of our work and your constructive suggestions.

Reviewer_7yGsOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses fine-grained cross-view localization and argues for using coarser localization priors and aerial image patches covering larger areas, such as 400 × 400 m. It shows that the performance of previous methods degrades quickly when the search area size increases. The paper proposes RingMatch, a search and verification framework that first identifies possible aerial image patches and then zoom into it for fine-grained localization. The proposed method surpasses fine-grained cross-view localization methods designed for smaller patches and a SOTA fine-grained method combined with cross-view image retrieval.

Critical strengths
  • It is an interesting finding that previous fine-grained cross-view localization methods degrade quickly with the increasing aerial patch size.

  • The proposed method takes reasonable design choice. Search and verification make sense to me.

  • The comparison also includes a baseline that combines a fine-grained method and retrieval.

Critical weaknesses
  • The caption of Fig. 1 says "existing fine-grained cross-view localization methods incur substantially larger errors as the aerial region expands from 35×35m to 400×400m". I presume 35×35m corresponds to the central area in the original vigor's aerial image patch (positive samples). In this case, does 400×400m mean the true location can be at the border of the image, or does the image cover an area larger than 400×400m? I feel there is some inconsistency between the size of the aerial image and localization prior.

  • The experiments are conducted on VIGOR only. I wonder how the depth ring-based "place recognition" generalizes to images with a limited field of view, and in general, how the proposed method performs on images with a limited field of view.

  • I find the idea of increasing the aerial image patch size a bit hand-wavy. I am not clear on the practical need for this.

  • Does the proposed method generalize to smaller aerial image sizes? Does search and zoom in help in that case?

Questions for authors

I don't have any further questions than the ones mentioned in weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer 7yGs (Part 1/2)

We sincerely thank the reviewer for the detailed evaluation and constructive comments. Below, we clarify the problem setting and provide further evidence and discussion to address the concerns raised.

1. Geographic prior in Fig. 1 (W1)

In VIGOR-Pro, the geographic prior is defined by the full area covered by the input aerial image. As the reviewer correctly noted, the conventional VIGOR protocol restricts the ground-camera location to the central 35 × 35 m region of a positive aerial tile. However, this center-biased assumption is restrictive in practice, since a coarse location estimate does not guarantee that the camera lies near the center of the aerial image. VIGOR-Pro therefore removes this restriction by expanding the aerial coverage and allowing the camera to appear throughout the image, including near its boundaries.

As detailed in Appendix B.1, the 400 m aerial images are constructed from 10 × 10 overlapping VIGOR tiles, with the source tile randomly positioned within the mosaic and the camera coordinates updated accordingly. Thus, 400 m refers to the aerial image coverage itself, not a localization region within an even larger image.

We agree that Fig. 1 may cause confusion between aerial image coverage and geographic prior. We will revise the axis label to “Geographic Prior Extent (m)” and clarify the corresponding descriptions throughout the paper.


2. Generalization beyond VIGOR and to limited ground FoVs (W2)

Our evaluation extends beyond the VIGOR benchmark. As reported in Table 4 of our submission, we additionally evaluate RingMatch on four cities outside the VIGOR benchmark—London, Tokyo, São Paulo, and Cape Town—without target-city adaptation. At the 200 m scale, RingMatch achieves a mean localization error of 13.09 m, compared with 16.77 m for Sample4Geo + FG², with consistent improvements across all four cities.

RingMatch can also be adapted to limited-FoV ground observations. A partial ground view produces a depth arc corresponding to the observed azimuth range, which can be aligned with the corresponding segment of each aerial depth ring. We evaluate 180° and 90° FoVs at the 200 m and 400 m scales under same-area evaluation.

Localization performance under different ground FoVs (mean / median error in meters).

ScaleMethod360°180°90°
400 mQwen3-VL-4B (End-to-End)23.03 / 3.9350.44 / 7.1378.99 / 17.10
RingMatch (Search-only)15.60 / 1.9236.34 / 3.3063.52 / 6.54
RingMatch12.36 / 1.8829.18 / 3.0955.98 / 5.64
200 mQwen3-VL-4B (End-to-End)12.16 / 3.5224.96 / 5.8538.71 / 10.13
RingMatch (Search-only)8.03 / 1.8217.14 / 3.0030.23 / 4.81
RingMatch6.57 / 1.7914.30 / 2.8826.31 / 4.41

Search and Verifier are trained separately for each FoV.

RingMatch remains effective under limited ground FoVs. Although localization errors increase as the visible region narrows, Search and Verification continue to provide complementary gains at both 180° and 90° FoVs. Even at 90° FoV, Verification reduces mean localization error from 63.52 m to 55.98 m at the 400 m scale and from 30.23 m to 26.31 m at the 200 m scale.

Verification accuracy (%) under different ground FoVs.

Scale360°180°90°
400 m82.8180.1371.79
200 m81.9278.0170.67

Depth-ring-grounded verification remains effective under limited FoVs. At 180° FoV, the Verifier maintains performance close to the full-view setting. Even at 90° FoV, it remains effective, achieving 71.79% and 70.67% verification accuracy at the 400 m and 200 m scales, respectively. These results demonstrate that depth-ring-grounded verification remains useful for distinguishing ambiguous locations even with a narrow ground FoV.

AuthorsOpenReview ↗

Response to Reviewer 7yGs (Part 2/2)

3. Practical need for a wide-area prior (W3)

Our motivation is the gap between coarse upstream localization and the narrow geographic priors assumed by existing methods. As discussed in our Introduction, GPS/GNSS or image retrieval may provide only coarse location estimates (Xia et al., 2022; Shi et al., 2023), which may not satisfy the 35 × 35 m geographic prior assumed in the conventional VIGOR protocol. This mismatch is further illustrated by reported localization errors in practical settings. For example, Lindenberger et al. (2025) report that 40.8% of queries have localization errors exceeding 100 m, including 31.3% exceeding 200 m, in continent-scale image geo-localization. Similarly, Hsu et al. (2023) report a 47.39 m RMSE and a maximum error of 209.25 m for smartphone-based 3DMA GNSS in harsh urban conditions.

These observations motivate extending fine-grained cross-view localization to broader geographic priors while maintaining high positioning accuracy. Accordingly, we construct VIGOR-Pro with geographic priors of approximately 100, 200, and 400 m to evaluate localization under increasing geographic uncertainty. We will clarify this practical motivation in the revised paper.

  1. Xia, Z., et al. (2022). Visual Cross-View Metric Localization with Dense Uncertainty Estimates. ECCV.
  2. Shi, Y., et al. (2023). Boosting 3-DoF Ground-to-Satellite Camera Localization Accuracy via Geometry-Guided Cross-View Transformer. ICCV.
  3. Lindenberger, P., et al. (2025). Scaling Image Geo-Localization to Continent Level. NeurIPS.
  4. Hsu, L.-T., et al. (2023). Hong Kong UrbanNav: An Open-Source Multisensory Dataset for Benchmarking Urban Navigation Algorithms. NAVIGATION, 70(4).

4. Performance on smaller aerial regions (W4)

RingMatch is specifically designed for wide-area localization under coarse geographic priors, rather than replacing specialized methods for tightly constrained local refinement. Nevertheless, to address this concern, we additionally evaluate RingMatch on the original VIGOR aerial tiles, including positive pairs with a 35 m geographic prior. Search uses a two-level branch-and-zoom schedule (70 m → 35 m).

Localization performance under smaller geographic priors (mean / median error in meters).

MethodALL (70 m, positive + semi-positive)Positive (35 m prior)
FG² (our reproduction)5.34 / 2.422.26 / 1.21
Qwen3-VL-4B (End-to-End)4.71 / 1.343.06 / 1.13
RingMatch (Search-only)4.43 / 1.182.67 / 1.07
RingMatch4.41 / 1.182.82 / 1.03

Search remains effective under smaller geographic priors. At 70 m, RingMatch achieves the lowest mean localization error (4.41 m), outperforming both FG² and direct localization. The branch-and-zoom process further refines localization, reducing median error from 1.22 m to 1.18 m after zooming, with comparable mean errors. Under the tighter 35 m prior, Search-only continues to outperform direct localization, although FG² achieves the lowest mean error.

Smaller geographic regions inherently involve less localization uncertainty. A tighter geographic prior already eliminates much of the location uncertainty, simplifying hypothesis exploration for Search and reducing the ambiguity that Verification needs to resolve. Consequently, the additional gains from both stages naturally become less pronounced, consistent with RingMatch's design for localization under substantial uncertainty.


We hope our clarifications and additional evidence help resolve the reviewer's concerns. We will include these improvements in the revised manuscript to better support our findings. We appreciate the reviewer's thoughtful suggestions.

I thank the authors for their rebuttal. After considering their response and the other reviews, I will maintain my initial rating.

Thank you for taking the time to review our rebuttal. We have provided detailed clarifications and additional experimental evidence in response to all four concerns raised in your review. We would appreciate it if you could clarify which specific technical concerns remain unresolved after considering our rebuttal, so that we can address them explicitly.

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

This paper investigates precise cross-view localization over substantially larger aerial regions than those considered in conventional local-refinement settings. To address the challenges of sparse cross-view correspondences and visually ambiguous locations, the authors propose RingMatch, a search-and-verification framework. The method first employs a multi-hypothesis branch-and-zoom strategy, using stochastic predictions from a fine-tuned vision-language model to identify plausible locations and progressively refine them across multiple spatial scales. It then introduces a depth-ring representation that encodes azimuth-wise distances to surrounding scene surfaces. Ground depth rings are derived from estimated monocular depth, while aerial depth rings are predicted from candidate-centered satellite patches. A fine-tuned LVLM verifier uses these geometric representations and visual context to select the final location. The authors also construct VIGOR-Pro by assembling existing VIGOR aerial tiles into larger contiguous regions. Experiments across multiple scales, geographic splits, and additional unseen cities demonstrate improvements over representative localization baselines.

Critical strengths
  1. A practically relevant problem setting with comprehensive empirical evaluation. The paper addresses an important limitation of existing fine-grained cross-view localization methods, which typically rely on narrowly constrained geographic priors. Extending precise localization to aerial regions of several hundred meters is a meaningful research direction. The proposed VIGOR-Pro benchmark provides a systematic evaluation of this setting. The experimental results demonstrate consistent improvements across multiple spatial scales and both same-area and cross-area settings. The additional evaluation on four unseen cities further strengthens the evidence for geographic generalization.

  2. A promising combination of multi-hypothesis search and explicit geometric verification. The search-and-verification framework is well motivated. Preserving multiple location hypotheses helps mitigate premature commitment during coarse-to-fine localization, while the depth-ring representation provides complementary geometric information for distinguishing visually similar locations. The ablation studies support the usefulness of both components. In particular, the depth-ring-grounded verifier improves localization compared with search alone, and the ring-only comparison demonstrates that the proposed representation contains discriminative spatial information.

Critical weaknesses
  1. The benefits of multi-hypothesis search are not sufficiently evaluated against its additional computational cost. The proposed branch-and-zoom strategy is effective, but its algorithmic novelty and practical efficiency require further clarification. Hierarchical coarse-to-fine localization has already been explored in related work, and the main distinction here lies in stochastic coordinate sampling, clustering, and preserving multiple continuous location hypotheses. Although the paper compares multi-hypothesis search with a single-hypothesis variant, the two approaches do not necessarily use comparable inference budgets. The proposed method can explore multiple branches at each level, resulting in more LVLM evaluations. Consequently, part of the improvement may be attributable to increased test-time computation rather than the specific uncertainty-guided branching mechanism. This concern is particularly relevant because the framework is motivated by practical localization under coarse geographic priors. However, the paper does not report the complete inference latency, average number of expanded nodes, or computational cost relative to competing methods. Since the method combines repeated LVLM predictions, depth estimation, aerial depth-ring prediction, and a separate verifier, its accuracy–efficiency trade-off remains unclear.

  2. The geometric validity of the depth-ring representation and its independent contribution require stronger evidence. The depth ring is presented as a view-shared geometric representation of distances to surrounding scene surfaces. However, ground rings are derived from monocular depth estimates rather than directly measured geometry, and the aerial ring predictor is trained using these estimated rings as supervision. The paper provides evidence that the representation is useful for localization, but limited quantitative evidence that it accurately captures the intended metric spatial structure. There is also a potential mismatch between the aerial predictor's input coverage and its depth prediction range. The predictor uses a 640 × 640 aerial crop while predicting distances up to 80 m. Given the reported ground sampling distances, a centered crop covers only approximately 30–40 m from the candidate position along cardinal directions. It is therefore unclear how reliably the predictor handles distances beyond the directly observable aerial context. Furthermore, the verifier ablation in Table 3 does not include a search-initialized verifier without depth-ring inputs. While the existing results support the usefulness of depth rings, this additional control would help establish their incremental contribution within the complete system.

Questions for authors

Q1. Can the authors clarify the computational cost and isolate the benefit of multi-hypothesis branching under a matched inference budget? Please report the average end-to-end inference latency, number of expanded search nodes, and computational resources required per query, particularly at the 400 m scale. Could the authors also compare against a single-path or multi-start search strategy with a comparable total number of model predictions? This would help determine whether the improvements arise primarily from the proposed branching mechanism or from additional test-time computation.

Q2. How accurately do the predicted depth rings represent the intended metric scene geometry? Could the authors provide a quantitative evaluation of aerial depth-ring prediction accuracy against the ground-derived supervision, preferably with an analysis across different distance ranges? In particular, how are directions handled when the target distance exceeds the spatial coverage of the input aerial patch? Clarifying the proportion of saturated or unobservable depth values would strengthen the geometric justification of the representation.

Q3. What is the independent contribution of depth-ring information within the complete verifier? Table 3 demonstrates improvements from depth-ring inputs and search-model initialization, but a search-initialized verifier without depth rings is missing. Could the authors include this control to isolate the contribution of the geometric representation? Additionally, given the strong performance of the ring-only correlation baseline, it would be informative to clarify the additional benefit of the learned LVLM verifier over direct geometric similarity-based candidate selection. Addressing these questions, particularly the efficiency comparison and the independent contribution of depth-ring verification, would substantially influence my overall assessment.

4 comments
AuthorsOpenReview ↗

Response to Reviewer TWqs (Part 1/2)

We sincerely thank the reviewer for the careful assessment and constructive suggestions. Below, we provide additional experimental results and clarifications to address the concerns raised.

1. Multi-hypothesis search under matched inference budgets (Q1)

To isolate the benefit of uncertainty-guided branching from additional inference computation, we compare RingMatch against Single-path and Multi-start search under matched coordinate sampling budgets. All strategies use the same Search checkpoint and zoom schedule.

  • Single-path: Retains only the centroid of the largest cluster at each level to define the next search region.
  • Multi-start (K=2 or 4): Divides root coordinate predictions into K equal groups to initialize independent search paths, each following the largest-cluster centroid at subsequent levels.

For a fair comparison, we record RingMatch's coordinate sampling budget for each query at each zoom level and apply the same budget to the baselines. Single-path concentrates samples on one region, whereas Multi-start distributes them across K independent paths.

Matched-budget search results under same-area evaluation at the 400 m scale.

Search strategyBudget / queryNodes / queryTop-1 mean (m) ↓Best@2 mean (m) ↓Best@5 mean (m) ↓
Single-path80.674.0018.0717.5917.47
Multi-start (K=2)80.677.0016.7814.0513.30
Multi-start (K=4)80.6713.0016.0813.1810.45
RingMatch (Search-only)80.675.0415.6010.498.53

Budget counts the stochastic coordinate samples requested from the Search model. Best@k mean error denotes the top-k oracle mean localization error, computed by averaging the minimum distance to the ground-truth location among the top k candidates across all queries.

Uncertainty-guided branching improves candidate quality. Although Multi-start (K=4) achieves comparable Top-1 performance, the primary objective of Search is to identify promising hypotheses for subsequent verification. RingMatch achieves a lower Best@2 mean error (10.49 vs. 13.18 m), with similar improvements in Best@5. These results demonstrate that uncertainty-guided branching better preserves promising hypotheses under matched coordinate sampling budgets.


2. End-to-end latency and inference cost (Q1)

We evaluate end-to-end inference latency and computational cost in the same-area setting at the 400 m scale. All methods run on a single NVIDIA A100 80 GB GPU, processing one query at a time after warm-up. Timing includes image preparation and all inference stages. For Sample4Geo + FG², the pipeline includes image preparation, encoding the ground image and all aerial tiles, retrieval, and FG² refinement. We also evaluate RingMatch with different search depths (L=2, 3, and 4) to demonstrate the flexibility of our method in balancing localization accuracy and inference efficiency.

End-to-end latency and localization error under same-area evaluation at the 400 m scale.

MethodLatency, mean / P95 (s)Localization error, mean / median (m)
Qwen3-VL-4B (End-to-End)0.964 / 1.01623.03 / 3.93
Sample4Geo + FG²2.149 / 2.70015.53 / 4.72
RingMatch (L=4)1.881 / 2.86012.36 / 1.88
RingMatch (L=3)1.639 / 2.46012.85 / 2.12
RingMatch (L=2)1.405 / 2.13914.25 / 2.83

Mean stage latency per query under same-area evaluation at the 400 m scale.

MethodImage prep. (ms)Search (ms)Depth-ring prep. (ms)Verification (ms)
RingMatch (L=4)542.91242.449.345.6
RingMatch (L=3)541.5994.352.150.6
RingMatch (L=2)542.0746.758.058.2

Image prep. includes ground and aerial image I/O and preprocessing. Depth-ring prep. includes ground depth estimation, aerial ring prediction, and rendering. Stage latencies are averaged over all queries.

Mean RingMatch workload per query under same-area evaluation at the 400 m scale.

Search depthSearch nodesCoordinate samplesGround ringsAerial ringsVerifier callsVerified queries (%)
L=45.0480.670.1370.2740.13713.7
L=33.5757.050.1520.3040.15215.2
L=22.1934.980.1570.3140.15715.7

RingMatch achieves a favorable and flexible accuracy–efficiency trade-off. Compared with Sample4Geo + FG², RingMatch (L=4) achieves 20.4% lower mean localization error and 12.5% lower mean latency, although its P95 latency is slightly higher. Search dominates inference time, while depth-ring preparation and verification add only 94.9 ms per query on average, with verification required for just 13.7% of queries. Reducing the search depth further decreases latency with a modest increase in localization error, allowing the trade-off to be adjusted to different application requirements.

AuthorsOpenReview ↗

Response to Reviewer TWqs (Part 2/2)

3. Depth-ring fidelity and aerial crop coverage (Q2)

As discussed in Sec. 3.3, depth rings serve as spatial signatures that capture azimuth-wise variations in surrounding scene layout, rather than requiring precise absolute-distance measurements. Since ground rings are derived from monocular depth estimates, we manually examine 100 ground–aerial correspondences to assess the relationship between DAV2-estimated depths and physical distances.

Manual comparison of physical distances and DAV2 depth estimates.

Actual distance (m)Actual mean (m)DAV2 mean (m)Overestimation factor
0–208.4032.483.87×
20–4027.1768.132.51×
>4043.3874.631.72×

The overestimation factor is the ratio of mean DAV2 depth to mean physical distance.

DAV2 exhibits substantial absolute-scale overestimation (1.72×–3.87×), suggesting that relatively large depth values may still correspond to scene structures within the 640 × 640 aerial crop. For structures that extend beyond the spatial coverage of the input aerial patch, the aerial-to-ring predictor is trained to infer the corresponding depth information from the available visual context. Detailed training procedures are provided in Appendix C.2.

To quantify prediction quality, we evaluate aerial depth rings against ground-derived supervision at the annotated camera locations under same-area evaluation at the 400 m scale.

Aerial depth-ring prediction quality across depth ranges.

Ground-derived depth (m)Azimuths (%)Median AE (m) ↓MAE (m) ↓Pearson w/ ground ring ↑
0–209.764.6510.94–
20–4031.554.067.92–
40–6024.497.1010.72–
60–8034.203.107.03–
Overall100.004.498.600.763

Pearson correlation is computed between predicted aerial and ground-derived depth rings for each query and then averaged across queries.

The predictor achieves a median absolute error of 4.49 m and MAE of 8.60 m. Among valid azimuths, only 3.82% of ground-derived depths and 4.12% of aerial predictions fall into the maximum depth bin (80 m). While image-based geometric estimation inevitably introduces point-wise errors, our objective is to preserve discriminative azimuth-wise spatial patterns rather than reconstruct precise scene geometry. The predicted and ground-derived rings achieve a Pearson correlation of 0.763, indicating good agreement in their overall spatial patterns. We have also provided qualitative depth-ring comparisons in Appendix F (Fig. 9) to illustrate this spatial consistency.


4. Independent contribution of depth rings and LVLM verification (Q3)

To isolate the independent contribution of depth rings, we additionally evaluate a Search-initialized Verifier without depth-ring inputs. Furthermore, to better assess the discriminative value of depth rings independently of learned LVLM verification, we introduce an additional probabilistic ring-matching baseline (NLL). NLL matching evaluates ground-ring depths under each candidate's predicted aerial depth distributions and selects the candidate with the lowest negative log-likelihood.

Verification ablation under same-area evaluation at the 400 m scale.

MethodSearch init.Depth ringALL mean (m) ↓VFY mean (m) ↓VFY acc. ↑
Pearson ring matchingN/AYes15.2367.8869.42%
NLL ring matchingN/AYes13.8557.8775.56%
Verifier w/o ringsYesNo14.5362.7974.44%
Verifier w/ ringsYesYes12.3647.1282.81%

ALL denotes all test queries, while VFY denotes the subset requiring verification.

Depth rings and learned LVLM verification provide complementary benefits. With Search initialization held fixed, adding depth rings improves verification accuracy from 74.44% to 82.81% and reduces localization error. The learned Verifier also outperforms direct geometric matching, achieving higher accuracy than both Pearson and NLL ring matching. Unlike direct geometric matching, which can be brittle under local depth errors, the learned Verifier jointly considers depth-ring patterns and visual context to more robustly distinguish ambiguous candidates.

5. Distinction from existing hierarchical coarse-to-fine methods (W1)

As discussed in Appendix A, prior methods such as Just Zoom In and GeoMoE primarily narrow geographic candidates through predefined discrete cells or gallery entries. In contrast, RingMatch performs uncertainty-guided multi-hypothesis search in continuous coordinate space: stochastic coordinate sampling reveals multiple plausible locations, clustering forms spatial hypotheses, and uncertainty-guided branching preserves and progressively refines them.


We hope these additional results and clarifications adequately address the reviewer's concerns. We will incorporate them into the revised manuscript to further strengthen our work. We sincerely thank the reviewer for the constructive feedback.

Thank you for the rebuttal and clarifications. I have reviewed the authors' responses and will maintain my original rating.

Thank you for taking the time to review our rebuttal and clarifications. We appreciate your thoughtful consideration and valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.