acceptodds
Under review as a conference paper at ICLR 2027

Test-Time Denoising for Noisy Cross-view Geo-localization

Abstract

Cross-view geo-localization (CVGL) aims to localize a ground-level query by retrieving its corresponding image from a geo-tagged satellite gallery. Although recent methods achieve strong performance on clean benchmarks, their accuracy can degrade when query images are affected by sensor, acquisition, or transmission noise. Existing noise-robustness solutions typically rely on training strategies or specialized model architectures, which require modifying or retraining the downstream localization system. In this work, we adopt a training-free and plug-and-play strategy from a shared denoising perspective. We distinguish noise-dependent observation statistics from the shared underlying clean-image target and prepend a frozen pretrained denoiser to existing CVGL models while keeping the localization model unchanged. The same pretrained denoiser is applied across heterogeneous query-side noise without noise-specific adaptation. Extensive experiments across multiple CVGL architectures, noise types, and severity levels show that pretrained denoising can improve localization robustness, with the gains becoming more pronounced under stronger noise. These results demonstrate the potential of pretrained denoising as a simple test-time strategy for improving the noise robustness of frozen CVGL models without retraining or modifying the localization network.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Cross-view geo-localization, Robust geo-localization, Image denoising

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_7uGPOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses noisy cross-view geo-localization (CVGL). In real-world settings, query ground images often suffer from various types of noise, while existing methods mostly assume that both query and reference images are clean. This work proposes a test-time plug-in method for existing CVGL methods to improve their robustness to noisy queries. Specifically, it uses a frozen DRUNet which is an existing image denoising network to denoise query images before feeding them into different CVGL networks. Experiments on CVACT-C show that adding this denoising module improves the performance of most evaluated methods.

Critical strengths
  1. The paper provides clear mathematical formulations which make its main idea easy to follow.
Critical weaknesses
  1. The novelty of this work does not appear sufficient for acceptance as a conference paper. Its core contribution is to place an existing image denoising network before an existing CVGL network. There is limited novelty in the architecture or model design, and the problem formulation itself does not appear sufficiently significant to compensate for this limitation.

  2. The writing includes too many formulas and details that are only loosely related to the method and do not help establish its effectiveness. For example, the paper describes four noise sources and their statistical properties. However, it uses a single denoising network to handle all four types without additional training, so these properties do not appear to inform the methodological design. The paper also contains too many defensive statements.

Questions for authors
  1. Could you compare the proposed method with recent approaches addressing noisy CVGL, such as MRGeo.
  2. Could you evaluate the performance of your proposed method on more recent state-of-the-art CVGL approaches, such as CV-Cities, and report the resulting gains of adding an image denoising network like DRUNet.
  3. DRUNet introduces substantial computational overhead while providing only a 4% gain. Please justify the practical value of this trade-off.
Reviewer_3zQ9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

Paper proposes test-time denosier to enhance noisy cross-view geo-localization (cvgl) task. Author proposes main problem in cvgl is during actual depoloymnet of cvgl, sensor noise, transmission noise, and other noise can signficantly affect the result. In prior works, special training strategies or model designs are required to achieve such task. However, paper proposes pretrained denoiser to denoise noisy query (usually street-view) to get clean prediction before cvgl with gnss-tagged satellite images.

Critical strengths
  1. Easy to follow: The problem set up is easy to follow where noise problem in computer vision is traditionally researched task when depolying to real-world scenarios from the research task.

  2. Test-time cvgl: Through pretrained or off-the shelf pretrained denoiser, cvgl task become plug-in-play task to adopt into real-world scenario.

Critical weaknesses
  1. Simple probblem formation: paper used gaussian, shot, speckle, impulse noise options. But these are simple simulated noise that does not covey real world noise problem. Noise probblem such as distortion of different types of cameras (low quality camera to high quality), weather condition, or objects blocking the sight are also reasonabbble noise in real-world scenarios rather than proposed simulation based noise.

  2. Novelty: I put easy to follow as strength but in other word, paper lacks in novelty. It used existing pretrained denoiser for noise removal and cvgl task. It seems to be more of engineering trick rather than academic reaserch task.

Questions for authors

see weakeness

  1. does author try truncated diffusion idea? using pretrained diffuson mode, instead starting with full gaussian noise, inputing noised query image to denoise. Since diffuions trained more larger number of images, this may performs well.
Reviewer_AKEGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The research scenario of the paper is relatively clear: ground query images are corrupted by noise, while the satellite gallery remains clean, and existing cross-view geo-localization (CVGL) models cannot be retrained. The processing pipeline adopted by the authors is denoising first, followed by matching. The experiments involve two datasets, seven localization models, four types of noise, and five severity levels. The main findings of the paper are: the average retrieval performance improves for most "model-dataset" combinations; the gains are substantial under heavy noise; useful discriminative information may be lost under light noise; and there are significant variations among different denoisers, localization models, and datasets. This work is best positioned as an "empirical study on transferring pre-trained denoisers to frozen CVGL systems." Its research value should primarily be supported by the reliability of the experiments, the depth of the analysis, and generalizable conclusions.

Critical strengths
  1. Specific problem and reasonable application constraints: The paper targets the scenario where "existing localization models are already deployed and difficult to retrain, but the quality of query images changes." Freezing the localization model, keeping the satellite gallery unchanged, and adding a processing step only on the query side is an easily understandable and highly reusable solution. The authors also clearly distinguish the statistical characteristics of different noises, rather than treating Gaussian, shot, speckle, and impulse noises as the same distribution. Furthermore, the main text on page 4 explicitly acknowledges that a shared clean target does not imply that the optimal denoiser is the same under different noise conditions. This is a prudent technical statement.
  2. Some performance gains are indeed substantial: This demonstrates that for certain frozen localization models, query-side denoising can rescue a considerable portion of retrieval failures caused by severe noise.
  3. Experimental coverage and disclosure of negative results are commendable: The evaluation across seven architectures, two datasets, four noise types, and five severity levels is much more comprehensive than a single-model, single-noise demonstration. The authors also report negative results for GeoDTR+, performance drops caused by DnCNN, computational overhead, and the limitation of unverified real-world noise. The appendix also provides feature similarity and retrieval gap analyses. Therefore, this paper cannot be simply dismissed as having "no analysis" or "only showcasing success cases." My main doubt lies in whether the existing analysis is sufficient to explain the differences among these results and to form insights that go beyond specific experimental configurations.
Critical weaknesses

W1. Weak methodological contribution; "shared clean target" is insufficient to support the main novelty. From Eq. (9) and Eq. (14), the actual algorithm simply places an existing denoiser in front of an existing localization model. The paper does not propose a new denoiser, a new training objective, new test-time optimization, or a new retrieval rule. The current "shared clean target" mainly describes the general structure of image restoration problems, a fact that cannot solely explain the cross-noise effectiveness of a fixed denoiser.

W2. Core denoising configuration is not disclosed, making main results difficult to reproduce. Eq. (9) emphasizes that γD\gamma_D is fixed, but neither the main text nor the appendix specifies what this fixed configuration actually is. Meanwhile, there is a lack of information regarding the noise-level map values and units for DRUNet; the noise level input for FFDNet; the specific pre-trained variant of DnCNN; the candidate range, selection criteria, and validation data for denoising parameters; whether the exact same configuration is used for both datasets and all seven localization models; the execution order of denoising, scaling, and normalization; and the denoising input resolution.

W3. Explicit numerical inconsistencies exist between the main tables and the appendix. For example, the averaged results of the graded evaluations in Tables 3–5 are inconsistent with the results in Table 2. The authors should regenerate the main tables, appendix, figures, and textual values from the same set of raw evaluation logs, rather than simply assuming one table is correct. If the appendix results are taken as the ground truth, some average gains and drops also need to be recalculated.

W4. The performance drop under light noise is not small, and a critical baseline comparing clean input passing through the same denoiser is missing. On page 8, the drop at the lowest severity level is described as "a small decrease," which is an understatement. According to Table 3, Sample4Geo drops by 4.17 percentage points on CVUSA-C at Severity 1. More critically, the paper completely lacks a localization performance comparison between: Clean queryandDθ(Clean query;γD)\text{Clean query} \quad \text{and} \quad D_\theta(\text{Clean query}; \gamma_D) The clean reference line in Figure 4 cannot replace this experiment. While equally weighting the four noise types and five severity levels is a reasonable baseline evaluation method, the proportions of clean, light-noise, and heavy-noise queries in actual deployment may not be the same. For a permanently enabled fixed frontend, the cost in clean and light-noise scenarios will directly determine the overall benefit. Therefore, I believe the following must be added:

  1. Results of all main models with clean input passing through the same denoiser;
  2. Overall performance under different proportions of clean, light-noise, and heavy-noise scenarios;
  3. The proportion of cases that were originally correct but became wrong after denoising, and cases that were originally wrong but were rescued by denoising;
  4. A sensitivity analysis of the fixed denoising strength. The failure case in the appendix shows the GT rank dropping from 5026 to 8147, both of which are already severe failures. In contrast, cases where the Raw rank = 1 but the denoised rank > 1 are much more revealing of the negative impact in actual usage.

W5. Ablation studies and mechanistic analysis are insufficient to explain "why it works and when it fails." The paper compares DRUNet, DnCNN, and FFDNet, but Figure 3 only provides aggregated results on Sample4Geo / CVUSA-C, indicating that the choice of denoiser is highly critical. This supports the claim that "DRUNet is effective under this setting," but the support for the "broad effectiveness of pre-trained denoising strategies" remains limited. I believe three targeted comparisons need to be added: 1. Simple preprocessing baselines; 2. More comprehensive cross-denoiser comparisons; 3. Background comparisons with robust CVGL methods.

W6. Significant computational overhead, lacking sufficient evidence for practical deployment value. In Figure 3, the computation of Sample4Geo increases from 61.1 GFLOPs to 545.7 GFLOPs, while the corresponding average R@1 increases from 81.76 to 86.16, a gain of about 4.4 percentage points. In heavy-noise scenarios, this cost might be worth bearing, but the paper has not yet provided sufficient information for readers to judge:

  • What hardware, input resolution, and batch size are used?
  • What is the actual latency per query?
  • How do peak VRAM and throughput change?
  • Do the FLOPs only calculate the online query, or do they include the satellite branch?
  • How does the accuracy-latency trade-off compare to cheaper preprocessing alternatives? Avoiding extra training is a clear advantage, but this cannot directly justify a low deployment cost. The authors have honestly reported the FLOPs, which is commendable; the next step is to support the claim of being "practical" with actual measurements. Furthermore, an 8.93× increase in FLOPs cannot be directly interpreted as an 8.93× increase in latency.
Questions for authors
  • What exactly is the fixed configuration? Please provide the noise condition inputs, checkpoints, preprocessing order for DRUNet/FFDNet, and the data source for parameter selection. Is the configuration unified across all datasets and localization models?
  • Which set of numerical values (main table or appendix) corresponds to the final experiments? Please explain the aforementioned cross-table inconsistencies and regenerate the results from a unified log.
  • What is the net benefit of always enabling the denoising frontend? How does the clean image perform after passing through the same denoiser? Under what deployment distribution can the light-noise loss be offset by the heavy-noise gain?
  • Compared to simple preprocessing, how much extra value does the pre-trained denoising prior bring? After selecting parameters on an independent validation set, how do median filtering, smoothing, or traditional denoisers perform?
  • Can the architecture and dataset dependencies be explained with verifiable evidence? Why does GeoDTR+ consistently suffer on CVACT while other combinations benefit? Is this difference mainly determined by the fixed denoising strength?
  • What is the actual online inference cost? Please report the per-query latency, VRAM usage, hardware specifications, and input size, and explain how the offline encoding of the satellite gallery is accounted for.
Reviewer_etsbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies cross-view geolocation under noisy query conditions and proposes adding a frozen pretrained image denoiser to the query side of existing frozen CVGL models while keeping the satellite gallery, retrieval rules, and localization network parameters unchanged. Motivated by the idea that different noises share a common clean-image restoration target, the authors mainly use DRUNet and additionally compare DnCNN and FFDNet. Experiments are conducted on CVUSA-C and CVACT val-C, covering seven CVGL models, four noise types, and five severity levels. The results show that this preprocessing strategy improves average R@1 in most model–dataset combinations and generally yields more pronounced gains under severe noise. However, its effectiveness depends significantly on the localization model, dataset, and noise severity, with degradation occurring in some settings. This paper presents an engineering baseline worth attention, but the current version still has substantive shortcomings in terms of contribution positioning, experimental controls and justification of deployment value.

Critical strengths

S1. The research question is clear, and the application constraints are practically meaningful.

For localization systems that cannot be retrained or are inconvenient to retrain, studying query-side preprocessing is reasonable. The method does not require modifying the localization network or rebuilding the satellite gallery, and it can be easily integrated into existing pipelines.

S2. The experiments have good cross-model coverage.

The evaluation covers seven CVGL models, two datasets, and four noise types at five severity levels. The results do not come from only a single model or a single noise setting.

S3. Some performance improvements are substantial in magnitude.

For example, in Table 2, SAFA on CVUSA-C improves average R@1 by 25.16 percentage points, and DSM improves by 21.67 and 14.97 percentage points on the two datasets, respectively. These results show that an independently pretrained denoiser can effectively improve the robustness of frozen CVGL models under some conditions.

Critical weaknesses

W1. The core conceptual contribution is limited, and there is a lack of substantive connection between the formal analysis and the method design.

Sec 3.2 states that different noises correspond to the same clean-image target and describes noise statistics through conditional mean and variance. This statement is basically reasonable, but it does not explain why the same fixed denoiser should be effective across noises, nor does it derive principles for denoiser selection, conditional parameter settings, or judging performance degradation. The authors have correctly noted that different noises do not necessarily share the same optimal denoising estimator. Therefore, “shared target” can currently only serve as motivation for empirical research and is not yet sufficient to constitute an independent theoretical or methodological contribution. Simple methods can have important value, but this paper needs more discriminative experimental findings to support this positioning.

W2. Key controls are missing, making it impossible to distinguish the contribution of the pretrained prior from the effect of general noise suppression.

The main experiments mainly compare original noisy inputs with DRUNet-preprocessed inputs. The lack of simple smoothing, median filtering, and other classical or low-cost denoising methods prevents readers from determining how much irreplaceable benefit the complex pretrained module actually brings. The comparison of DnCNN and FFDNet is only provided as aggregated results on Sample4Geo/CVUSA-C, which is insufficient to support conclusions about denoiser selection across models and datasets. For a paper whose main contribution is empirical, these controls are more important than further adding results under similar settings.

W3. Key implementation and selection protocols are not sufficiently disclosed, which affects reproducibility.

Equation (9) introduces a fixed conditional parameter γD\gamma_D, but the paper does not report the actual conditional values used by DRUNet and FFDNet, their units, or the basis for their selection. The specific checkpoint, DnCNN version, order of denoising and resizing, input/output ranges, and padding strategy are also insufficiently clear. In addition, the authors state that DRUNet was selected as the default method because of its highest empirical robustness, but they do not clarify which data were used for model and configuration selection and whether these data are independent of the final reported sets. This does not directly prove that test-set tuning occurred, but it needs clarification.

W4. The harm under mild noise cannot be ignored, and there is a lack of evaluation of clean inputs passed through denoising.

Sec 4.5 describes the degradation under the mildest noise as small, but in Table 3, Sample4Geo/CVUSA-C at Severity 1 drops from 97.43% to 93.26% in average R@1; in Table 5, the corresponding drops for GeoDTR+ on the two datasets are 6.61 and 6.37 percentage points, respectively. These results are directly relevant to a deployment strategy that enables preprocessing by default without identifying the noise state. The clean-input baseline in Figure 4 cannot replace the experiment of “clean inputs passed through the same denoiser.” The absence of this control prevents the paper from fully assessing the risk of fixed preprocessing for normal queries.

W5. The computational cost is high, and the practicality conclusion needs to be narrowed or supported with additional evidence.

In Figure 3, DRUNet increases the reported total computation from 61.1 to 545.7 GFLOPs, about 8.93 times the original, while the corresponding average R@1 improves by about 4.39 percentage points. The larger gains under severe noise may justify this cost, but the paper does not provide actual latency, throughput, peak memory usage, or a clear basis for online query computation. Therefore, "no retraining required" and "easy to integrate" are supported, but it cannot yet be directly inferred that the deployment cost is acceptable.

W6. The result analysis lacks uncertainty and sufficiently in-depth explanations of heterogeneity.

The paper does not report confidence intervals, nor does it state whether multiple independent noise realizations were evaluated. For small changes such as +0.44 percentage points on L2LTR/CVACT, the stability of the results is difficult to judge. The feature analysis in Appendix C is a useful supplement, but the retrieval margin lacks a clear mathematical definition, and the analysis covers only one model–dataset combination. It has not yet explained why the same denoising treatment produces significantly different effects on different models. The equally weighted average over five severity levels can serve as a benchmark metric, but it should not be directly equated with the expected benefit under real deployment distributions.

Questions for authors

Q1. Fixed configuration and model selection.

What are the exact conditioning parameters used for DRUNet and FFDNet? How were these values determined, and are they kept consistent across models and datasets? Were the denoiser and configuration selections made using the final reported sets?

Q2. Competitiveness of simple baselines.

How much additional benefit does DRUNet retain compared with median filtering, simple smoothing, and classical denoising methods? Do the conclusions still hold under a comparable computational budget?

Q3. Normal queries and mild-noise risk.

How does the performance of each CVGL model change when the same fixed denoising pipeline is applied to clean queries? At the per-query level, how many originally correct top-1 results are destroyed, and how many originally incorrect top-1 results are recovered?

Q4. Sources of model dependence.

Why does GeoDTR+ degrade on average on CVACT val-C while other models benefit substantially? Is this difference more likely due to architecture, input processing, fixed denoising strength, or different reliance on image details? What direct evidence is currently available?

Q5. Actual runtime cost.

What input size and scope do the FLOPs correspond to? Is satellite gallery precomputation separated from the online query cost? Under specified hardware and batch size, what are the single-query latency and peak memory usage after adding DRUNet?

Q6. Result stability and analysis definitions.

What are the paired confidence intervals for the main performance differences? Were different random noise realizations tested? In Appendix C, what negative-sample definition and aggregation method are used for the retrieval margin?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.