acceptodds
Under review as a conference paper at ICLR 2027

Beyond Invariance: Learning the Geographic Core for Generalizable Cross-View Geo-Localization

Abstract

Cross-view geo-localization (CVGL) identifies locations by matching images across different viewpoints. Existing methods prioritize view-invariant representations, but invariant semantics may exhibit high visual similarity across different locations, limiting geographic discrimination under domain shifts. Although scene semantics can narrow plausible matches, they cannot fully resolve ambiguous places, motivating complementary evidence from fine-grained regional structure. Inspired by sufficient invariant learning, we propose GeoCore to learn a geographic core of complementary semantic and structural evidence. The semantic branch can capture view-invariant scene semantics for coarse discrimination. To resolve the spatial ambiguity, the structural branch learns fine-grained regional structure that remains discriminative across environmental variations. Extensive experiments on UAV and ground-view benchmarks demonstrate that our model consistently outperforms existing approaches, achieving superior performance in cross-area generalization and robustness evaluations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Generalizable Cross-View Geo-Localization, Geographic Core, Sufficient Invariant Learning

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

Cross-view geo-localisation (CVGL) has attracted considerable attention due to its research value in cross-modal feature matching and its practical applications in visual localisation and vehicle navigation.

This paper addresses a key limitation of existing generalisable CVGL methods , namely, relying primarily on view-invariant visual features may lead to ambiguous matching when different locations share similar visual semantics. The authors propose GeoCore, a framework that combines semantic and structural cues to improve cross-domain localisation. GeoCore consists of two complementary branches. The Textually Anchored Semantic Core branch aligns image-level features with textual descriptions generated offline by a vision-language model (VLM) to learn transferable scene semantics. The Reliability-Aware Structural Learning branch aggregates spatial features into anchored regions, models their spatial arrangements through rotation-symmetric structural refinement, and estimates regional reliability under appearance variations.

The authors evaluated GeoCore on five UAV- and ground-view localisation benchmarks and reports improvements over existing methods, with strong results in cross-dataset generalisation.The resulting descriptor is trained using cross-view alignment and auxiliary semantic and reliability objectives.

Critical strengths
  1. Well-motivated problem formulation: Combining transferable scene semantics with fine-grained spatial structure is sensible. The two branches are complementary. Semantic features provide coarse geographic discrimination, while structural features help distinguish locations with similar scene-level semantics. I find Fig-5 which provides CAM visualisations of these complementary cues is convincing.

  2. Explicit modelling of regional reliability: Estimating the geographic discriminability of local regions under appearance variations is useful. Rather than simply aggregating local features, the method aims to identify regions that remain discriminative across different environmental conditions. This is a promising direction for improving robustness in cross-view localisation.

Critical weaknesses

--- . Limited technical novelty: Combining semantic and structural information is interesting, but the individual components, including CLIP-based semantic alignment, regional aggregation, 4-rotation-symmetry based invariances, and reliability prediction, are mostly established techniques. The paper does not clearly demonstrate what is technically novel beyond combining them in one framework. As shown in Figure 1, the design also involves considerable engineering designs and multiple components, which obscures the core idea. The authors should clarify the key insight that distinguishes GeoCore from existing methods.

---. The dependence on VLM-generated descriptions in training is unclear: The semantic branch relies on offline-generated textual descriptions during training, but their quality and consistency are not sufficiently discussed. The authors should examine in inference how sensitive the method is to description quality, different VLMs, and dataset-specific cues. They should also clarify how much improvement comes from textual supervision rather than the visual backbone itself.

Questions for authors
  1. What invariance does the four-rotation in RSSR provide? Can it handle arbitrary rotations, such as roads that are not aligned with the four 90-degree orientations?

  2. Textual supervision: How sensitive is GeoCore to noisy or incomplete VLM-generated descriptions? Do the results remain consistent when using different VLM models or simpler than CLIP descriptions? If possible within ICLR rule, can the authors provide additional experiments to show this ?

  3. There are some related relevant references missing: e.g.

  • From Satellite to Ground: Satellite Assisted Visual Localization with Cross-view Semantic Matching (ICRA 2024).
  • VGA: Empowering Aerial-Ground Localization by Visual Geometry Alignment (CVPR 2026).
  • Optimal Feature Transport for Cross-View Image Geo-Localization_ (AAAI 2020).
4 comments
AuthorsOpenReview ↗

Response to Reviewer gpCP (1/4)

We sincerely appreciate your valuable and constructive suggestions. Our point-by-point responses are provided below.


W1: Clarifications on Novelty

Most existing CVGL methods emphasize cross-view invariance. However, Fig. 1 reveals two limitations: invariant cues differ in geographic discriminability, and semantically similar locations can still cause ambiguous matches. This indicates that view consistency alone is insufficient, motivating complementary geographic evidence that is both transferable and discriminative.

GeoCore addresses the challenges through complementary semantic and structural learning. The semantic branch supplements visual cues with textual guidance to capture view-shared and transferable scene-level semantics. To resolve residual ambiguity among semantically similar places, the structural branch extracts fine-grained discriminative spatial features. SAR organizes spatial tokens into spatially anchored regions through content-spatial affinity, preserving regional relationships. Then, RSSR applies a shared nonlinear readout to four quarter-turn grid permutations and averages the outputs, reducing grid-orientation sensitivity while retaining spatial arrangements. And CRR promotes geographic discrimination across environments through training-only regional reliability supervision based on positive–hard-negative margins, without introducing additional modules during inference. Compared with prior methods, this branch learns more explicitly structured and orientation-robust regional representations. Overall, GeoCore learns transferable and geographically discriminative representations, supporting both UAV and Ground-view generalizable CVGL scenarios.

The experiments in Table 5 and Table 12 verify the complementary contributions of the two branches, while Table 11 and Reviewer HfHm's Table R2 support the effectiveness of RSSR and discrimination-aware CRR supervision. Importantly, Table 3 highlights the efficiency of image-only retrieval, demonstrating GeoCore’s practical applicability and deployment potential.

AuthorsOpenReview ↗

Response to Reviewer gpCP (2/4)

W2/Q2: Effectiveness of textual supervision

1. The quality control of the text-generation pipeline: Appendix A.2.4 describes our pipeline of text-generation quality control, which integrates cross-view verification and retrieval-oriented prompting. Rather than independently captioning images, we generate location-level descriptions from paired views, retaining shared scene elements and coarse spatial relations while filtering unsupported or view-specific cues (e.g., illumination). A consistent three-field format further standardizes the generated descriptions across different datasets. As clarified in 3.3, text is generated and encoded offline using a frozen CLIP text encoder, serving only as training supervision. GeoCore performs image-only retrieval at inference, without requiring extra textual descriptions.


2. Sensitivity analysis on textual descriptions: To evaluate sensitivity to imperfect textual descriptions, we introduce controlled text perturbations involving noise and incompleteness at different levels. In Table R1, both noisy and incomplete descriptions reduce image-text relevance (Column Sim: 0.6531 for University-1652→SUES-200 and 0.6241 for CVACT→CVUSA with the original descriptions), with stronger degradation under noise. Specifically, Sim decreases from 0.6531 to 0.4641 under 90% noise and to 0.5299 under 90% dropping on University → SUES, with consistent trends on CVACT → CVUSA. Despite the reduced quality, GeoCore maintains strong retrieval performance, achieving over 93% R@1 and 94% AP across all perturbation settings on University → SUES. These results demonstrate GeoCore's robustness to noisy and incomplete textual anchors, indicating that partially correct semantic information remains effective for learning geographic discrimination and supporting its practical applicability in real-world scenarios.

Table R1: Ablation results (%) on the qualities of perturbatant text anchor

MethodUniversity-1652 → SUES-200 (R@1 / AP / Sim)CVACT → CVUSA (R@1 / R@1% / Sim)
Noisy 50%94.71 / 95.65 / 0.568768.73 / 97.75 / 0.5140
Noisy 90%93.25 / 94.26 / 0.464168.28 / 97.58 / 0.4291
Drop 50%95.45 / 96.24 / 0.607468.66 / 97.87 / 0.5698
Drop 90%93.54 / 94.49 / 0.529968.09 / 97.46 / 0.5028

Notes on Table R1: (1) Perturbation protocol: We randomly select 50% or 90% of semantics (e.g., elements in Appendix A.2.4) across training locations and either remove the selected textual items (Drop/Incomplete) or replace them with randomly sampled items from other locations (Noisy). For Drop, at least one semantic item is retained per location. (2) Sim: The score measures cross-view image-text consistency. We use Qwen3-VL-Reranker-2B to assess the relevance of each description to its paired UAV and satellite images, computing the score as σ(logit_yes − logit_no) and averaging it across both views and all locations.


3. Dependence of text supervision: In HfHm's Table R2, removing Lsem\mathcal{L}_{\mathrm{sem}} reduces R@1 from 96.61% to 92.25% on University → SUES and from 72.60% to 70.43% on CVACT → CVUSA for GeoCore, confirming that textual guidance can promote semantic–structural complementarity and geographic discrimination. Furthermore, we evaluate alternative VLMs in Table R2. GeoCore consistently maintains competitive retrieval performance across different VLMs, including the smaller-scale Qwen3-VL-2B, demonstrating the effectiveness and transferability of textual supervision without relying on a specific VLM generator.

Table R2: Ablations results (%) on different VLMs for text anchor

VLMUniversity-1652 → SUES-200 (R@1 / AP)CVACT → CVUSA (R@1 / R@1%)
Qwen3-VL-2B95.91 / 96.6470.80 / 98.05
DeepSeek-V4.196.51 / 97.0972.70 / 98.18

The additional results and corresponding analysis are now included in the supplementary material (Appendix A.3.7(a), Table 14, page 30, and Tables 15-16, page 31).

AuthorsOpenReview ↗

Response to Reviewer gpCP (3/4)

Q1: More analyses on RSSR

RSSR applies a shared nonlinear readout to four quarter-turn permutations of the encoded regional grid and averages the outputs. As established in Appendix A.1.2 (Lemma 3), RSSR ensures invariance to discrete quarter-turn permutations of the regional grid, rather than arbitrary input-image rotations. Importantly, these permutations do not constrain road orientations: roads at arbitrary angles remain representable through local visual features and their spatial arrangements, without requiring alignment to the four canonical directions.

As shown in Appendix Table 11, RSSR improves R@1 over ordered readout (without rotation averaging) by 0.69 and 0.25 points on University → SUES and CVACT → CVUSA, respectively. Conversely, shuffling regional arrangements (permuting region features before RSSR) reduces R@1 by 1.47 and 1.18 points. Overall, Lemma 3 and these ablations support that RSSR reduces sensitivity to grid orientation while preserving spatial organization for geographic discrimination.

AuthorsOpenReview ↗

Response to Reviewer gpCP (4/4)

Q3: More relevant works

We thank the reviewer for the helpful suggestion and have incorporated these studies into our discussion in the revised manuscript (Related Work, page 2).

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

To address the challenge of cross-domain generalization in cross-view geo-localization, this paper proposes the GeoCore framework. It improves retrieval performance across different geographic regions and environmental conditions by combining scene semantics with fine-grained regional structures. The method has two branches: text-anchored semantic core and reliability-aware structural learning. The former uses location-level text descriptions generated by a vision-language model to provide semantic supervision. The latter learns structural representations through Spatially Anchored Regionalization, Rotation-Symmetric Structural Refinement (RSSR), and auxiliary supervision through Cross-Environment Regional Reliability (CRR). Features from both branches are concatenated to form a retrieval descriptor, and inference requires only an input image. The paper conducts cross-dataset evaluations on five UAV and ground-view datasets and reports strong cross-domain generalization and robustness to various weather conditions.

Critical strengths
  1. The paper studies a useful problem and proposes a clear method. The paper focuses on cross-domain generalization in CVGL. It combines scene semantics and regional structures to reduce confusion between similar locations. The two branches have clear roles, and the theoretical analysis and ablation studies provide support for the method.

  2. The experiments are extensive and show strong results. The paper evaluates GeoCore on five UAV and ground-view datasets, including different altitudes, weather conditions, and DINO backbones. The results show good cross-domain generalization. The paper also analyzes the contributions of different branches and modules, while keeping text generation and reliability prediction out of inference.

Critical weaknesses
  1. The baseline settings and the independent contributions of some modules are not clear enough. Tables 5 and 13 show improvements over DINO baselines, but the baseline architectures, descriptor designs, and training settings are not fully explained. It is unclear whether these baselines use the same optimization conditions as GeoCore. Also, the current ablations do not separately show the benefit of VLM text supervision or compare CRR with simpler auxiliary supervision. These details are important for understanding where the improvements come from.

  2. Some baseline results differ from the original papers, which may affect the reported performance gains. In Table 2, EP-BEV reports 59.32% and 32.68% R@1 on CVACT-val/test, while its original paper reports 67.79% and 44.10%. Also, the InfoGeo results mainly use the version without Relational Distillation (RD), and some weather results differ from its original paper. These differences may come from different implementations or evaluation settings. The authors should clarify the result sources and model versions, and update the comparisons if necessary.

Questions for authors

Q1. Baselines using the same backbone
Please clarify how the baselines in Tables 5 and 13 construct their descriptors and what training objectives they use. Which optimization settings are shared with GeoCore, including the learning rates, freezing strategies, and training budgets in Appendix A.2.2? Please also clarify whether they use the same source-domain splits and checkpoint selection criteria. If there are differences, how might they affect the comparisons?

Q2. Reported baseline results
Please provide the exact source of the EP-BEV results in Table 2 and explain the differences from Table 4 of arXiv:2408.05475v1. Please also specify the InfoGeo version and result sources in Tables 1, 3, and 8, and explain why the RD version was not included. If the comparison results need to be updated, please also correct the relative performance gains.

Q3. Complementarity, text supervision, and CRR
On one representative task, could the authors report the errors corrected and new errors introduced by fusion, using the branch descriptors from the same trained GeoCore model? An analysis of semantically similar hard negatives would also help. Could the authors also keep the semantic branch and remove only the text alignment loss to test the benefit of the VLM-generated text supervision? For CRR, could the authors compare soft-min and mean-margin targets while keeping the environmental variations and predictor unchanged?

3 comments
AuthorsOpenReview ↗

Response to Reviewer HfHm (1/3)

Thank you very much for your positive comments on our work and for the valuable suggestions. We respond to your points one by one as follows.


W1/Q1: Details on Baselines in Table 5 & 13

To assess the GeoCore's gains beyond pretrained backbones, we establish CLS-only baselines using the same DINOv2/DINOv3 backbones and optimizing with cross-view contrastive learning (InfoNCE), excluding both proposed semantic and structural branches. Both baselines follow the same training and optimization protocols, source-domain splits, and checkpoint selection criteria as GeoCore (Details in Appendix A.2.2), while only global [CLS] tokens are used for retrieval. Furthermore, Table 13 and additional results in Reviewer GiCU's Table R1/2 demonstrate consistent improvements over Baselines across transfer and multi-weather settings, confirming the additional generalization benefits of GeoCore across different visual backbones.

As shown in Table R2, our additional text-supervision ablation shows that incorporating textual anchoring by LsemL_{sem} alongside direct image alignment improves R@1 from 91.85% to 96.61% and AP from 93.08% to 97.20%, supporting semantic–structural complementarity. Furthermore, we compare CRR with a simpler supervision--Feature Consistency (FC), which averages cosine similarities between clean and perturbed regional features within each view, without cross-view correspondence or negative samples. FC reduces R@1 by 1.42/1.37% relative to GeoCore on University→SUES and CVACT→CVUSA, respectively. This result validates the benefits of geographic discrimination margins and cross-environment aggregation.


We have clarified the baseline configurations in Section 4.3 (Page 8) of the revised manuscript, with the revisions highlighted in blue.

AuthorsOpenReview ↗

Response to Reviewer HfHm (2/3)

W2/Q2: Reported Comparsion Results

For comparison methods with publicly available implementations, including EP-BEV, we reproduce their results using official code on a unified computation platform, following their respective training configurations and evaluation protocols.

As shown in Table 1, GeoCore achieves an average R@1 improvement of 4.26% over InfoGeo with Relational Distillation (InfoGeo+RD) across four UAV cross-dataset transfer settings. Under three adverse-weather conditions (Table 3), GeoCore and GeoCore$^\dagger$ further achieve average R@1 gains of 2.40 and 2.98 points, with corresponding AP gains of 2.44 and 2.70 points, respectively. These results further demonstrate the effectiveness of the proposed method in cross-domain generalization and robustness under challenging environmental conditions.


We have added the InfoGeo+RD results and corresponding analysis to the revised manuscript (Table 1, page 6; Table 3, page 7).

AuthorsOpenReview ↗

Response to Reviewer HfHm (3/3)

W1/Q3: More Ablations

In response to the reviewers' comments, we conduct additional ablation studies to further examine semantic-structural complementarity, textual supervision, and the individual contributions of CRR, as follows:

1. Error Correction Analysis: To quantify semantic–structural complementarity, we compare the fused descriptor with each branch descriptor from the same trained GeoCore model. As shown in Table R1, fusion consistently corrects more errors than it introduces. On University → SUES, fusion yields 1,151 and 969 net corrections (Corrected Errors - New Errors) relative to the Semantic and Structural Cores, respectively, while the corresponding numbers on CVACT → CVUSA are 2,545 and 82. These results demonstrate the complementary contributions of scene-level semantics and fine-grained regional structures to geographic discrimination across different scenarios.

Table R1: Error Correction Analysis of Semantic–Structural Fusion

Transfer SettingCoreCorrected Errors ↑New Errors ↓Net Corrections ↑
University → SUESSemantic1,198471,151
University → SUESStructural1,109140969
CVACT → CVUSASemantic2,6891442,545
CVACT → CVUSAStructural24115982

2. Effect of LsemL_{sem}: We retain the semantic branch while removing Lsem\mathcal{L}_{\mathrm{sem}} to examine the contribution of textual supervision. As shown in Table R2, R@1 drops by 4.36 and 2.17 % on University → SUES and CVACT → CVUSA with the DINOv2 backbone, respectively, confirming the benefit of the Textually Anchored Semantic Core. Consistent with the ablations in Table 5, these results highlight the complementary roles of text-anchored semantics and structural learning in improving generalization performance.


3. Ablations on LregL_{reg}: We also compare CRR with different supervision variants. While Feature Consistency (FC) encourages within-view consistency, Mean-margin averages geographic discrimination margins across environments. As illustrated in Table R2, the proposed Soft-min improves R@1 over Mean-margin by 0.70/1.04% with GeoCore and 0.53/0.26% with GeoCore$^\dagger$ on University→SUES and CVACT→CVUSA, respectively, with consistent gains in AP and R@1%. These results highlight the advantage of Soft-min in emphasizing challenging environments and preserving geographic discrimination.

Table R2: Additional ablation results (%) on textual supervision and regional reliability

MethodUniversity → SUES (R@1 / AP)CVACT → CVUSA (R@1 / R@1%)
GeoCore (w/o LsemL_{sem})92.25 / 93.4070.43 / 98.04
GeoCore (w/ LregL_{reg}, FC)95.19 / 96.0171.23 / 98.06
GeoCore (w/ LregL_{reg}, Mean-margin)95.91 / 96.5871.56 / 98.08
GeoCore$^\dagger$ (w/o LsemL_{sem})96.76 / 97.3870.78 / 98.09
GeoCore$^\dagger$ (w/ LregL_{reg}, FC)97.12 / 97.8974.90 / 98.28
GeoCore$^\dagger$ (w/ LregL_{reg}, Mean-margin)97.54 / 98.0175.11 / 98.32

The textual supervision and CRR ablations have been incorporated into the revised manuscript (Appendix A.3.7, Table 16, page 31).

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

This paper proposed the GeoCore, generating the geographic-core representation for CVGL task that aims to be both transferable across domains and discriminative between locations. It combines a text-anchored semantic core for transferable scene concepts with an image-based structural core for fine-grained regional information that can distinguish semantically similar places. The paper backs this design with extensive experiments: ablations support the utility of its components, and GeoCore outperforms the listed prior methods across several transfer and weather settings.

Critical strengths
  • Localizing in new settings without target-domain adaptation is an important practical challenge in CVGL. This work proposed an intuitive and effective approach, with extensive experiments reporting improvements over the listed prior methods across UAV and ground-to-satellite transfers and in the tested weather conditions.
Critical weaknesses
  • It is hard to tell how much the proposed design adds beyond the Baseline across the reported settings. This matters especially because the paper's central aim is to learn features that are both transferable and geographically discriminative. For example, on CVACT→CVUSA, the DINOv2 Baseline reaches 68.55 R@1, already 23.60 points above the best listed prior result of 44.95; GeoCore reaches 72.60, adding another 4.05 points. This gain is meaningful. But, baseline results are only reported for limited settings in the ablation section. Matched baseline results in the other settings with large reported gains are needed to assess how consistently and significantly the proposed design adds transferable discrimination, and whether that increment warrants its complexity.
Questions for authors
  • Could you describe in detail what the baseline in tables 5 and 13 are? Please also provide matched Baseline results for the remaining transfer directions and weather settings, especially those with the largest improvements over prior work.
  • What does the text anchor add to the semantic core beyond direct image-to-image alignment? how would retrieval and the learned semantic content change if sqs^q and sgs^g were aligned directly instead of, or alongside, the text-mediated loss under otherwise matched conditions?
2 comments
AuthorsOpenReview ↗

Response to Reviewer GiCU (1/2)

We appreciate the time and effort you have dedicated to our manuscript. Below, we summarize the key responses to your concerns.


W1/Q1: More Analyses on Baseline in Tables 5 & 13

1. Details on Baseline: The baselines in Tables 5 and 13 use the [CLS] tokens from DINOv2/DINOv3 under the same training protocols as GeoCore, excluding both proposed branches and retaining only cross-view InfoNCE alignment. Building on this baseline, GeoCore learns complementary scene semantics and fine-grained regional structures to improve geographic discrimination and cross-domain generalization. Importantly, GeoCore is compatible with different visual backbones and consistently improves their cross-domain generalization performance.


2. More experiments of Baseline: As shown in Table 13, GeoCore improves R@1 by 5.33/2.44 points with DINOv2/DINOv3 backbones on University-1652 →\to SUES-200, respectively, and by 4.05/5.97 points on CVACT →\to CVUSA. We further conduct matched-baseline experiments on the remaining transfer and weather settings in Tables R1 and R2. DINOv2 baselines generally trail DINOv2-based methods (e.g., InfoGeo/CVcities) in many scenarios. Meanwhile, GeoCore consistently improves over matched baselines, with R@1 gains of 5.50--18.70% across transfer settings and 7.96--11.82% under adverse weather. With GeoCore$^{\dagger}$, the corresponding gains reach 4.96--23.14% and 4.81--6.66%, respectively. These consistent improvements demonstrate the effectiveness and transferability of our design across different visual backbones.

Table R1: More experimental results (%) on the generalization scenarios

MethodUniversity → DenseUAV (R@1 / AP)DenseUAV → SUES (R@1 / AP)DenseUAV → University (R@1 / AP)CVUSA → CVACT-val (R@1 / R@1%)CVUSA → CVACT-test (R@1 / R@1%)
DINOv2 Baseline36.34 / 28.0889.40 / 90.6565.34 / 66.1862.63 / 96.2829.22 / 43.79
DINOv3 Baseline38.44 / 28.7594.23 / 95.2780.84 / 83.7162.88 / 97.3031.22 / 45.73

Table R2: More experimental results (%) on the multi-weather evaluation

MethodFog-Rain (R@1 / AP)Fog-Snow (R@1 / AP)Rain-Snow (R@1 / AP)
DINOv2 Baseline74.36 / 78.8274.01 / 78.0384.77 / 87.20
DINOv3 Baseline80.03 / 82.4377.42 / 80.1588.53 / 89.66

3. Complexity Analysis: We also analyze computational complexity and find that GeoCore introduces only a modest gain in network parameters (86.88M to 99.61M) and inference latency (12.05 to 12.59 ms) compared with the DINOv2 Baseline. Compared with the DINOv2 Baseline, GeoCore modestly increases parameters (86.88M to 99.61M) and inference latency (12.05 to 12.59 ms), while delivering substantial generalization gains. Since CVGL has limited real-time demands and satellite-gallery features can be precomputed offline, retrieval accuracy remains a priority. In this context, the modest computational overhead is justified by the substantial R@1 improvements, demonstrating a favorable accuracy-efficiency trade-off.


The details of the baseline have been clarified in Section 4.3 (Page 8) of the revised manuscript, with the corresponding revisions highlighted in blue.

AuthorsOpenReview ↗

Response to Reviewer GiCU (2/2)

W2/Q2: More analyses of Text Anchor

As illustrated in Fig. 1, conventional image-to-image alignment primarily emphasizes cross-view consistency, without explicitly distinguishing the geographic discriminability of different visual cues. Consequently, semantically similar locations may remain difficult to distinguish. To address this limitation, GeoCore introduces view-consistent, location-level textual descriptions as semantic anchors, providing explicit guidance on shared scene concepts and coarse spatial relations. This encourages the semantic branch to learn transferable geographic semantics beyond visual correspondence alone.

1. Ablation on text anchoring: To explore the effectiveness of text anchoring, we replace the text-mediated loss with direct image-to-image alignment under the same training settings. As shown in Table R3, direct semantic alignment reduces R@1 by 4.76/2.10% on University → SUES and 6.12/5.35% on CVACT → CVUSA for GeoCore and GeoCore$^\dagger$, respectively. This is further corroborated by Reviewer HfHm's Table R2, where removing Lsem\mathcal{L}_{\mathrm{sem}} also leads to consistent R@1 degradation across both transfer settings. These results demonstrate the significant benefit of textual supervision beyond direct visual correspondence learning.

Table R3: Ablation results (%) on direct image-to-image alignment

MethodUniversity → SUES (R@1 / AP)CVACT → CVUSA (R@1 / R@1%)
GeoCore91.85−4.7691.85_{\scriptscriptstyle -4.76} / 93.08−4.1293.08_{\scriptscriptstyle -4.12}66.48−6.1266.48_{\scriptscriptstyle -6.12} / 96.52−1.6896.52_{\scriptscriptstyle -1.68}
GeoCore$^\dagger$95.97−2.1095.97_{\scriptscriptstyle -2.10} / 96.91−1.5496.91_{\scriptscriptstyle -1.54}70.02−5.3570.02_{\scriptscriptstyle -5.35} / 97.70−0.6397.70_{\scriptscriptstyle -0.63}

2. Impact on semantic content: Regarding the learned semantic content (sqs^q, sgs^g), compared with textual anchoring, direct image-to-image alignment on GeoCore improves the Semantic Core's retrieval performance (R@1) from 44.08% to 62.37% (+18.29%) on CVACT → CVUSA and from 89.96% to 90.96% (+1.00%) on University → SUES, respectively, but reduces Structural Core R@1 from 71.81% to 66.86% (-4.95%) and from 92.78% to 72.13% (-20.65%), respectively. These findings suggest that direct image alignment encourages the Semantic Core to absorb additional visual matching cues, potentially weakening its complementarity with structural learning. In contrast, textual anchoring promotes semantic-structural complementarity by guiding the Semantic Core toward transferable scene semantics, while allowing the Structural Core to capture fine-grained regional distinctions that resolve semantic ambiguity. This is consistent with our motivation in Fig. 1: cross-view invariance alone does not guarantee geographic discrimination.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Thank you to all the reviewers for their constructive comments

Summary of Rebuttal Responses and Major Manuscript Revisions

We sincerely thank all reviewers for their constructive feedback. We have conducted additional experiments and revised the manuscript to address the key concerns. The rebuttal responses are summarized below:

  • Clarified technical novelty (gpCP). We highlight that cross-view invariance alone is insufficient for geographic discrimination, particularly when visually similar scenes correspond to different locations. GeoCore addresses this limitation by learning complementary semantic and structural evidence to capture transferable scene-level semantics and extracts fine-grained discriminative structural information.

  • Expanded baseline evaluations (GiCU, HfHm). We clarify the CLS-only DINOv2/DINOv3 baselines and extend matched comparisons across transfer and multi-weather settings, confirming consistent gains with both backbones. We analyze computational complexity, demonstrating efficient inference and a favorable accuracy–efficiency trade-off. Manuscript Revisions: Baseline configurations are further clarified in Section 4.3 (Page 8).

  • Semantic–structural complementarity (HfHm). We introduce error-correction analysis to quantify the complementary contributions of both branches, demonstrating that semantic and structural cues provide complementary evidence for geographic discrimination.

  • Text-anchor robustness (GiCU, gpCP). We introduce controlled noisy and incomplete text perturbations, evaluate alternative VLMs (Qwen3-VL-2B and DeepSeek-V4.1), and compare textual anchoring with direct image-to-image alignment, demonstrating robustness to imperfect descriptions and different VLMs while validating the additional benefit of textual supervision. Manuscript Revisions: Additional evaluations are included in Appendix A.3.7(a), Table 14 (Page 30) and Tables 15–16 (Page 31).

  • More ablations on structural branch (HfHm, gpCP). We clarify RSSR's discrete quarter-turn invariance and use ordered-readout and region-shuffling experiments to validate the ability to reduce grid-orientation sensitivity while preserving spatial organization. We further evaluate different CRR variants, demonstrating the benefit in maintaining geographic discrimination and regional reliability across environmental variations. Manuscript Revisions: Supporting results are detailed in Appendix Table 16.

  • Updated comparisons and related work (HfHm, gpCP). We incorporate stronger comparisons, including InfoGeo+RD, reproduce publicly available codes and expand discussions of relevant studies. Manuscript Revisions: Updated comparisons appear in Tables 1 and 3 (Pages 6–7), with additional discussions in Related Work (Page 2).

Overall, we have carefully addressed all reviewers' comments above and would welcome any further questions or discussion.

Thank you for your encouraging feedback and positive assessment of our additional experiments. We are pleased that the new results and clarifications have addressed your concerns, particularly regarding the consistent improvements over matched baselines and the benefits of textual semantic supervision. We also appreciate your insightful suggestion on exploring visual reasoning alone as an alternative means of extracting view-shared semantic information for CVGL. We have included more future work on this direction in Appendix A.5 Discussion, Page 34.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.