acceptodds
Under review as a conference paper at ICLR 2027

Fog-Robust Cross-View Geo-Localization: Datasets and Framework

Abstract

Cross-view geo-localization (CVGL) aims to match ground-view queries with satellite images, yet existing works largely assume clear visibility and rarely consider adverse weather conditions such as fog. Fog usually obscures textures and structural cues essential for cross-view correspondence. To enable controlled investigation, we construct two synthetic foggy CVGL datasets, CVUSA-Fog and CVACT-Fog, using the atmospheric scattering model while preserving the original ground–satellite correspondence. Beyond synthetic evaluation, we manually curate 1,441 real-fog ground-view images and pair them with satellite images to construct CVFog. Additionally, we further propose a plug-and-play Wavelet-guided Attenuation Structural Adapter (WASA) module, which couples low-frequency degradation cues with structural responses to selectively calibrate fog-affected yet informative regions in ground-view features. Experiments across multiple backbones, datasets, and fog intensities show that synthetic-fog training substantially improves both fog-domain retrieval and synthetic-to-real performance on CVFog, while WASA provides consistent further gains. We further study semi-supervised real-fog adaptation with limited pairing supervision. Under the same real-fog pairing budget, confidence-filtered hard pseudo-pairs improve retrieval over supervised adaptation by exploiting pair-unlabeled CVFog images.

Primary area:
datasets and benchmarks
Keywords:
Cross-view geo-localization, Fog robustness, Fog datasets, Structure-aware feature calibration, Semi-supervised adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_dwYhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

THe paper investigates how to identify the location of a ground-level photograph by matching it against satellite images, even when fog covers the scene. In order to do this the authors create artifical fog versions of two existing datasets, and collet a new set of real and fog image pairs. They also add a component to prior matching models in order to help them use visible edges and structures despite the fog.

Their experiments show that training with the aritificial fog improves matching on real photographs with fog, and that the proposed component provides additional improvements across several models. Their main contribution is the new data for evaluation and incremenetal improvements on previously existing methods.

Critical strengths

The real fog evaluation makes the work very interesting practically. Related similar approaches like MRGeo have only tested on artificially corrupted fog.

WASA's benefits are quiet consistent. Results are reported across a decent number of backbones and datasets with good ablation.

Critical weaknesses

The paper doesn't compare against MRGeo. This is a recent method that's designed to handle degradations including fog. Without a comparison it's difficult to judge how much the proposed method gains over existing work.

The ablation experiments testing which parts of the system are import only use the artificial fog. That makes it unclear whether the full system is needed for real photographs with fog or whether a less complicated system would have worked as well.

Questions for authors

It would be good to know the query set sizes for tables 2 and 4. And also the averageing procedure. Some percentages (e.g. 17.26% R@1) don't seem to fit the pattern, can these numbers even be generated from 1441 queries?

How was the spatial separation and duplication removal implemented for the datasets? Geographic distances/overlap thresholds/etc?

As per above, how does MRGeo perform when trained an evaluated in the same procedure. I guess its existing benchmark scores are not eauivalent to those here.

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

This manuscript introduces two synthetic fog datasets, CVUSA-Fog and CVACT-Fog, and a real-fog dataset, CVFog, for ground-to-satellite geo-localization. The authors also propose WASA, which combines wavelet-based low-frequency statistics with Sobel-guided structural responses to calibrate ground-view features, and further evaluates semi-supervised real-fog adaptation using confidence-filtered pseudo-pairs.

Critical strengths

The experiments shows the benefit of their proposed training pipeline. In that section, they separated the benefit of synthetic-fog training from the additional contribution of WASA and show improvements across multiple backbones and real-fog images. Comparisons with mixed-intensity training, generic adapters, and learnable filters, together with cross-dataset evaluations, provide useful evidence of the method’s effectiveness under different conditions. Furthermore, more importantly, they construct high-quality of dataset for society.

Critical weaknesses

I think the experiment fully shows the effectiveness of WASA, which is meaning that the fog-specific behavior of WASA needs further validation. The low-frequency energy ratio can also be high in naturally smooth regions, and Tab. 3 and Fig. 9 do not isolate whether the improvement comes from fog-specific calibration, or not. I think that the authors should analyze matched clear and fog responses against the known synthetic transmittance and compare with the same gate without the energy statistic.

Secondly, the contribution of each pseudo-pair filtering condition remains unclear. Tab. 7 shows that the complete filtering procedure improves retrieval, but adding the margin condition alone reduces R1, leaving it unclear whether both margin and consistency conditions are needed. A consistency-only variant and a random-selection control retaining the same number of pseudo-pairs would help isolate the benefit of the proposed combined rule.

Lastly, the real-fog evaluation needs clearer settings and closer comparisons. Tab. 2 does not specify the synthetic training source or the exact CVFog query and gallery configuration, while the criteria for removing spatially overlapping samples remain qualitative. These settings should be reported explicitly, and the cited MRGeo should be evaluated under the same data and training conditions to establish the advantage over an existing method designed for corruption robustness.

Questions for authors

Please address all of the weakness points. I think the biggest contribution is from the constructed dataset. If authors addressed all the concerns well, this paper can be accepted.

Reviewer_e2f6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies cross-view geo-localization (CVGL) under fog, which prior work largely ignores since it assumes clear visibility. For this specific setting, the author (i) first introduces two synthetic fog datasets, CVUSA-Fog and CVACT-Fog, built with the atmospheric scattering model from manually screened and repartitioned ground-view pairs, with fog applied only to the ground view so correspondence is preserved, plus CVFog, comprising 1,441 manually curated real-fog pairs. (ii) It then proposes WASA, a plug-and-play adapter placed after the first stage of the ground-view encoder, where a wavelet branch estimates low-frequency degradation, a fixed-Sobel branch extracts remaining structure, and their interaction gates a residual calibration toward fogged yet structurally informative regions. (iii) It further studies semi-supervised real-fog adaptation using confidence-filtered hard pseudo-pairs from unpaired images. The experiments spans four backbones on both synthetic-fog datasets, synthetic-to-real tests on CVFog, and ablations on components, filters, fog intensity, and alternative modules, mostly on Sample4Geo.

Critical strengths
  1. A comprehensive and detailed plan was designed to create a fog enhanced CVGL dataset for this issue, providing sufficient infrastructure for the research community.

  2. To address this issue, the model can adaptively enhance its understanding of downstream task related regions by introducing wavelet transform and Sobel operator prior features.

  3. A comprehensive experimental analysis was conducted, from overall results to ablation experiment analysis, and the robustness of the model under different data strategies was analyzed, enhancing the credibility of this paper.

Critical weaknesses

I have doubts about the value of this issue. The main model for this task can achieve over 95% performance. Moreover, the main model base is directly tested under foggy conditions, which mainly affects R1. In other words, the range of influence during direct foggy testing is mainly limited to the few candidates mentioned earlier. So I am skeptical whether CVGL under foggy conditions is sufficient to support the definition of a new problem.

I am curious if a defogging model can already solve this problem. The paper also discussed this issue and compared the Dehazing baseline with the WASA module in Figure 7. But there are a few questions: firstly, whether adding a dehazing model to this setting is already sufficient. Because the Dehazing baseline has a small difference in performance compared to WASA, with latency already below 5ms, I think it is already good enough. Secondly, is the Dehazing baseline training free or training? Can the total parameters be controlled to the same level and experimentally tested if it is training? If it is training free, is a dehaze baseline already good enough because it is more convenient and likely to generalize better without relying on specific training data? Can the preliminary results of this baseline dehaze be presented in the paper and visualized as a case study to observe whether dehaze itself can obtain a sufficiently good preliminary result. Thirdly, it is necessary to demonstrate the difference in generalization between the two and conduct a cross test set evaluation comparison.

There is another small issue, the CVUSA chart in Figure 1 appears to be misplaced.

Questions for authors

Do we necessarily need to provide this' manually fogged 'dataset for the CVGL task of solving this fog setting? The method in the paper is essentially a data augmentation method that calls the deep model atmospheric scattering model. I would like to see what results can be achieved by training the baseline using more relevant data augmentation methods, such as masks? Can this setting also be solved relatively well.

RealFog SSL Hard uses a significant amount of pseudo pairing compared to Real Sup. This is essentially 'using more data' rather than 'better self supervised methods'. Lack of random pseudo pairing/additional synthetic fog pairing with the same amount of data as a control.

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

This paper studies ground–satellite cross-view geo-localization under foggy conditions. The authors construct two synthetic fog datasets, CVUSA-Fog and CVACT-Fog, from CVUSA and CVACT, and additionally collect, manually curate, and pair a real fog dataset, CVFog, comprising 1,441 image pairs from Mapillary. Methodologically, the paper proposes WASA, which calibrates ground-view features through low-frequency feature statistics, structural responses, and gated residual updates. The authors further study semi-supervised adaptation under limited real paired supervision, using consistency and similarity-margin filtering of teacher-generated hard pseudo-pairs.

The work targets a practically meaningful problem, and the complementary design of synthetic and real data, the integrable feature-adaptation module, and the relatively broad experimental coverage are valuable. However, the manuscript does not yet sufficiently establish its incremental contribution relative to existing adverse-weather localization studies, nor does it adequately disentangle the sources of performance gain between WASA and the semi-supervised pipeline. Several details regarding data splits, evaluation protocols, and label budgets also require further clarification.

Critical strengths

The research problem is specific and has practical deployment significance. The impact of fog on ground-view images is asymmetric relative to the imaging conditions of satellite reference images, which further exacerbates cross-view matching difficulty. The paper targets preserving location-discriminative cues and improving retrieval representations, rather than assuming that visually clear images must first be restored. This is a reasonable task-specific focus.

The complementary design of synthetic and real data has potential resource value. CVUSA-Fog and CVACT-Fog support controlled experiments on fog intensity, while CVFog is used to examine synthetic-to-real transfer. The two types of data serve different roles, which is more valuable than evaluating only within a single synthetic degradation pipeline. In particular, the collection and ground–satellite pairing of real fog images in CVFog require substantial practical data-curation effort. If its pairing quality, spatial splits, and availability are ensured, this dataset has the potential to become a useful resource for future research.

WASA has sound task motivation and integrability. WASA attempts to combine degradation-related cues with still-observable structural information to calibrate retrieval features, and the overall design motivation is coherent. The module is inserted only into the ground encoder, requires no additional degradation or structural annotations, and follows the original backbone’s retrieval objective, giving it practical potential for adaptation to existing CVGL systems.

Critical weaknesses

W1. The incremental contribution over existing adverse-weather CVGL research is insufficiently established, and comparisons with the closest prior work are missing.

The paper addresses an important problem, but its positioning does not sufficiently distinguish its contributions from existing adverse-weather and corruption-robust CVGL research. In particular, Benchmarking the Robustness of Cross-view Geo-localization Models (ECCV 2024) already introduced CVUSA/CVACT corruption benchmarks containing fog at multiple severity levels and investigated augmentation-based robustness improvements, yet this directly relevant benchmark is neither cited nor compared against. MuSe-Net also studies weather-dependent feature adaptation and includes synthetic adverse-weather experiments on street-to-satellite CVUSA, not only drone-to-satellite datasets. Thus, the ground–satellite setting alone does not distinguish the present work from existing weather-aware localization research.

At the methodological level, MRGeo directly addresses corrupted-image street-to-satellite CVGL through dynamically gated feature fusion, channel calibration, and geometric structural constraints. Although the manuscript cites MRGeo, it provides neither an experimental comparison nor a sufficiently detailed account of how WASA differs from this closely related approach. Comparisons with CBAM and Conv Adapter support WASA’s effectiveness relative to generic alternatives, but do not establish its competitiveness against methods specifically designed for robust CVGL.

The potentially meaningful distinctions, depth-dependent fog synthesis, a curated real-fog dataset, and a lightweight fog-oriented adapter, should therefore be established more explicitly. A particularly informative evaluation would compare the proposed synthesis pipeline with existing fog corruption/augmentation under matched training conditions and assess transfer to CVFog. The authors should also compare against the most relevant adaptable robust CVGL method under the same data splits and supervision, or clearly explain the technical constraints preventing a meaningful comparison.

W2. The experiments establish WASA’s utility more clearly than its proposed mechanism of degradation-aware structural calibration.

Combining established components is not inherently a weakness. The issue is whether the experiments isolate the capability introduced by the combination of Haar DWT, low-frequency statistics, Sobel responses, gating, and residual calibration. The results support the conclusion that WASA improves retrieval under the reported settings, but they provide less direct evidence that it specifically identifies and calibrates regions that are both fog-degraded and geographically discriminative.

First, the manuscript acknowledges that high low-frequency energy can also occur in naturally smooth regions. However, it does not directly demonstrate that the learned degradation cue distinguishes fog attenuation from scene smoothness or other forms of detail loss. The low-frequency haze prior itself has precedent in WDMamba, but transferring an image-domain observation to a learned shallow-feature statistic requires validation in the present setting. Second, Table 3 primarily ablates the inputs to the calibration gate: even the “DC Only” variant still uses the structural candidate SgS_g in the residual update. These results therefore should not be interpreted as independently isolating the full degradation and structural branches. Third, the heatmaps and response–retrieval-margin correlations are consistent with the proposed explanation, but do not rule out general structural enhancement or other learned attention effects; correlations pooled across fog intensities may also reflect severity as a common factor.

The generic-module comparisons and fixed-Sobel-versus-learnable-filter ablation provide useful supporting evidence, but do not fully resolve these alternatives. More diagnostic tests would include controlled clear/fog comparisons of the same scenes, analysis of degradation responses against known synthetic attenuation while controlling for scene content, and capacity-matched low-pass or generic-gating alternatives. Without such evidence, the mechanistic interpretation should be presented as a plausible explanation rather than an established account of the gains.

W3. The source of the semi-supervised gains and the scope of real-world generalization require further clarification.

  • Semi-supervised adaptation.

Table 6 reports retained pseudo-pair precision of only 21.37%–22.82%. The authors appropriately acknowledge this substantial residual noise, so the concern is not undisclosed label unreliability, but the insufficiently explained usefulness of the retained supervision. In Table 7, adding weak-view consistency to margin filtering changes average precision from 22.12% to 21.68%, while increasing final R@1 from 30.67% to 34.75%. This does not contradict the reported improvement, but it indicates that pair correctness alone does not explain the gain. The contribution of stable sample selection, pseudo-pair identity information, and additional regularization from real-fog images remains unclear. Coverage-matched random selection, a consistency-only filtering variant, or an unpaired real-image consistency baseline would help distinguish these effects.

  • The relationship to existing semi-supervised CVGL should also be clarified.

UCVGL (CVPR 2024) already studies ground–satellite pseudo-pair learning using mutual matching and Top-1–Top-2 similarity-margin filtering, yet is not discussed or compared against. The manuscript currently frames this component as a further study of real-fog adaptation, which is reasonable; its contribution should be stated accordingly. If methodological novelty is intended, the authors should identify and substantiate the advance over these existing strategies, preferably through a matched-budget comparison. The CVFog results and cross-dataset experiments provide useful evidence beyond within-dataset synthetic-fog evaluation. Nevertheless, both synthetic datasets use the same fog-generation pipeline, so cross-intensity and cross-dataset gains alone cannot separate robustness to fog more generally from adaptation to that generator’s characteristic degradation. CVFog offers a meaningful real-world check, but the breadth of that evidence is difficult to assess without clearer reporting of geographic coverage, route/sequence composition, scene categories, and ground-view field of view. Although the authors describe removing near-duplicates and overlapping satellite crops, the operational criteria and resulting split statistics are insufficiently specified to assess spatial independence.

Questions for authors

Q1. What is the incremental value of the proposed datasets and method relative to the closest prior work?

Please explicitly compare CVUSA-Fog, CVACT-Fog, and CVFog with existing corruption and multi-environment benchmarks in terms of view setting, fog generation, real-fog coverage, and spatial partitioning. Which research questions do the proposed datasets enable that existing resources cannot readily address? Since MRGeo is already cited, can the authors provide a comparison under matched training and evaluation conditions, or explain the specific comparability constraints?

Q2. What evidence isolates fog-specific calibration from general feature enhancement in WASA?

Can the degradation response be validated against synthetic attenuation or controlled changes in fog intensity while accounting for naturally smooth scene regions? Please report WASA’s gain under clear-train/clear-test conditions. For Table 9, can a Sobel-initialized but trainable variant be included to distinguish the benefit of gradient initialization from that of keeping the prior fixed?

Q3. What explains the semi-supervised improvement despite substantial pseudo-pair noise?

Are incorrect pseudo-pairs visually similar or geographically close to their ground-truth matches, and how are cases in which multiple queries select the same satellite handled in the contrastive loss? Can coverage-matched selection controls or an unpaired consistency baseline distinguish the value of pseudo-pair identity information from that of stable sample selection and additional real-image regularization? How does the approach compare with existing CVGL pseudo-pair filtering strategies, particularly mutual matching and similarity-margin filtering?

Q4. How are spatial independence and reproducibility of the data splits ensured?

Please provide exact post-cleaning split sizes, sequence/spatial grouping rules, near-duplicate and satellite-overlap thresholds, and cross-split geographic distance statistics. For the synthetic datasets, how does repartitioning relate to the original CVUSA/CVACT splits, and were initialization checkpoints trained on any locations subsequently assigned to evaluation? Will the split identifiers, pairing metadata, cleaning criteria, and evaluation protocol be released?

Q5. Are evaluation conditions and supervision budgets fully comparable across experiments?

For each main table, please specify the source training dataset, evaluated split, query/gallery sizes, number of runs, and averaging procedure. Can Table 4 include a source-only Sample4Geo+WASA baseline evaluated on the identical CVFog test split and gallery, so that adaptation gains can be separated from differences in evaluation conditions? Finally, does the stated 5%/10%/20% pairing budget count only training correspondences? Please separately report the labeled validation data used for model selection and threshold calibration, including the implications of pooling predictions across label budgets.

Reviewer_8N4uOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates fog-robust CVGL. The authors construct two synthetic fog datasets, CVUSA-Fog and CVACT-Fog, using a depth-aware atmospheric scattering model, and introduce CVFog, a real-fog benchmark containing 1,441 ground–satellite pairs. To improve localization under foggy conditions, the paper proposes a Wavelet-guided Attenuation Structural Adapter (WASA), which combines low-frequency degradation cues and structural responses to enhance ground-view features. Extensive experiments evaluate the proposed framework across multiple CVGL backbones, fog intensities, and real-world fog conditions. The authors also explore semi-supervised adaptation using confidence-filtered pseudo-pairs under limited real-fog supervision.

Critical strengths
  • The paper addresses the degradation of cross-view geo-localization performance under foggy conditions, which is a practically relevant challenge for real-world localization systems.
Critical weaknesses
  • Insufficient novelty and inadequate discussion of prior work. The authors claim that existing CVGL studies largely overlook adverse weather conditions such as fog. However, this claim understates substantial prior research. Zhang and Zhu (ECCV 2024) [1] have already established corruption robustness benchmarks on CVUSA and CVACT, explicitly including fog at multiple severity levels. Furthermore, MRGeo (AAAI 2026) [2] and WMGR-Net (Pattern Recognition 2026) [3] have investigated degradation-aware feature enhancement and weather-robust representations. RHO (CVPR 2026) [4] has also studied fog-robust geo-localization through synthetic weather augmentation, albeit in an OSM-based metric localization setting. Although some related works are cited, the paper fails to adequately distinguish its contributions from existing studies and lacks comparisons with representative weather-robust methods. Consequently, the novelty of the proposed synthetic datasets and WASA framework remains insufficiently justified.

  • Insufficient validation of dataset realism and evaluation reliability. The validity of the constructed fog datasets is not sufficiently established. Since fog synthesis relies on per-image normalized monocular depth, the same attenuation coefficient does not necessarily correspond to consistent physical visibility across scenes. More importantly, Figure 10 presents only illustrative image-statistic distributions rather than empirical evidence demonstrating the similarity between synthetic and real fog. The real-fog dataset is manually selected from approximately 150,000 candidates, yet details regarding geographic diversity, fog severity, and potential selection bias remain limited. Although the authors mention removing near-duplicate images and overlapping satellite crops, more explicit criteria and reproducible split protocols are needed to ensure reliable evaluation.

  • Low pseudo-pair accuracy and unclear benefits of semi-supervised adaptation. The semi-supervised adaptation results raise serious concerns about pseudo-pair reliability. According to Table 6, only 21–23% of the retained pseudo-pairs are correct, despite a coverage of approximately 76–79%. Thus, the vast majority of retained pseudo-pairs provide incorrect supervision, making the mechanism behind the reported improvements unclear. Furthermore, the margin-only variant performs worse than the supervised baseline (Table 7), while the improvement under the 5% labeling budget is merely 0.72 percentage points. The authors should explain why such noisy pseudo-pairs improve retrieval and provide comparisons with alternative unlabeled-data learning strategies, such as consistency regularization, along with stronger statistical evidence.

References

[1] Zhang, Q., & Zhu, Y. (2024). Benchmarking the robustness of cross-view geo-localization models. European Conference on Computer Vision (ECCV), 36–53.

[2] Wu, L., Bo, L., Ouyang, S., & Zhu, Y. (2026). MRGeo: Robust cross-view geo-localization of corrupted images via spatial and channel feature enhancement. Proceedings of the AAAI Conference on Artificial Intelligence, 40(13), 10691–10699.

[3] Xu, S., Sun, J., Long, X., Liu, J., & Zhang, D. (2026). Weather-aware multi-granularity representation learning for cross-view geo-localization under adverse conditions. Pattern Recognition, 180, 114445.

[4] Zheng, J., Dai, R., Liu, R., et al. (2026). RHO: Robust holistic OSM-based metric cross-view geo-localization. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 33727–33737.

Questions for authors
  1. How do CVUSA-Fog and CVACT-Fog fundamentally differ from existing corruption benchmarks such as CVUSA-C, and why are representative weather-robust CVGL methods not included in the comparisons?

  2. Can the authors demonstrate that the proposed wavelet-guided calibration is more effective than simpler frequency filtering, edge-based enhancement, or parameter-matched attention modules?

  3. Is the relatively small improvement from the explicit interaction term (97.31% to 97.93% R@1 in Table 3) statistically significant across multiple random seeds?

  4. Why does replacing fixed Sobel operators with learnable filters lead to such a substantial performance drop (6.34 percentage points in Table 9), and could this be attributed to optimization instability?

  5. Why is GeoDTR excluded from the zero-shot evaluation on CVFog, despite being included in the synthetic-fog experiments?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.