acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Cross-View Geo-Localization: A Probabilistic Motivation

Abstract

Cross-view geo-localization (CVGL) matches a street-view query, typically a panorama, to a geo-tagged satellite reference image of the same location. Existing methods primarily address the severe viewpoint gap between the two modalities, but pay much less attention to appearance variation caused by weather and daylight. We model a CVGL image as , where is an image formation function and denotes location-preserving content, viewpoint, appearance, and residual nuisance. This formulation clarifies the learning objective: preserve information about while reducing sensitivity to the remaining factors. This probabilistic view motivates ProbInfoNCE, which replaces the shared temperature in InfoNCE with a learned pair-dependent scale to account for sample-dependent ambiguity during training. We also introduce CAV-Geo, a large-scale global dataset with appearance variation, with 256K street-view locations and 440K satellite-view images. CAV-Geo further includes a cross-continental split designed to evaluate out-of-distribution generalization. It is the largest CVGL benchmark with explicit weather and daylight variation, and its cross-continental protocol provides a substantially more challenging evaluation setting than existing standard splits. With the same DINOv3 initialization, ProbInfoNCE improves R@1 over InfoNCE by on VIGOR cross-area and on CVACTCVUSA. Appearance augmentation with dual-matrix training and query-to-query alignment further improves R@1 on original VIGOR cross-area queries by relative to ProbInfoNCE trained without appearance augmentation. All percentage gains in this paper are computed as ratios to their stated baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Geolocalization, Computer Vision

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

This paper proposes ProbInfoNCE, which uses a shared MLP to predict scaling components for query and reference images separately. These are then added to form a pairwise temperature. Under a conditional independence assumption, the analysis decomposes the expected squared distance between embeddings of the same location into query-side variation, reference-side variation, and cross-view mismatch. This provides motivation for the loss design. The authors also use synthetic appearance augmentation, dual-matrix training, and query-to-query alignment. They combine VIGOR and DRESS to construct CAV-Geo, which provides weather and day-night variants and a cross-continent protocol with training outside North America and testing in North America. The paper reports performance on several standard benchmarks and the new dataset.

Critical strengths

1.The method is simple and efficient.
ProbInfoNCE adds only a small MLP with little extra computation. Controlled experiments show that it improves R@1 from 64.74% to 68.07% on VIGOR Cross-Area with the same DINOv3 initialization.

2.The new benchmark is useful.
CAV-Geo covers 11 cities and 255,875 query locations, with synthetic weather and day-night changes. The authors also provide several checks to evaluate whether image editing preserves localization cues.

Critical weaknesses

1.The algorithmic novelty is limited.
Sample-dependent temperatures have already been studied. Table 8 shows improvement over query-side temperature scaling, but it is unclear whether the gain comes from adding reference-side information or using the additive form. More controlled comparisons are needed to support the specific contribution.

2.The robustness and generalization evidence is insufficient.
Both training and testing use FLUX-generated appearance changes, so performance on real weather or other distributions remains unclear. Also, CAV-Geo lacks a matched InfoNCE baseline to isolate the contribution of the new loss.

Questions for authors

Q1 Contribution of adaptive temperature
Could the authors add reference-side-only scaling and a few other ways to combine the query and reference scores to separate the effect of reference-side information from that of the additive form? Please explain the specific differences from Zhang et al. (2021) and Qiu et al. (2023), and why these differences matter.

Q2 External validity of appearance robustness
Could the authors evaluate the method on queries edited by a different generator, on an independent degradation benchmark, or on a small set of images captured in adverse weather? Cross-generator experiments could start with a representative subset.

Q3 Controlled experiments on CAV-Geo
In the Cross-Continent setting, could the authors match initialization, training data, and optimization settings, add both Dual Matrix and Dual Matrix + q2q configurations for InfoNCE, and report results for original and synthetic queries separately?

Q4 Stability of cross-continent results
Could the authors keep the full test reference gallery unchanged and provide R@1 for each of the five North American test cities? If results for other transfer directions are available, please also provide them.

7 comments
AuthorsOpenReview ↗

Response to Reviewer aTtE

We thank the reviewer for constructive feedback.

Scoring definitions. Retrieval tables report R@1 (%), and edited means FLUX unless specified. For L2-normalized query and reference embeddings qi,rjq_i,r_j, cosine ranks satellite candidates by qi⊤rjq_i^\top r_j. Native ranks them by ProbInfoNCE’s learned additive score (submission Eq. 11 and Appendix C, Eq. 32):

\qquad \epsilon=10^{-6}.$$ The two positive scales are predicted by the shared softplus MLP from the corresponding **pre-normalization features**. **Prob + cosine and Prob + native use the same ProbInfoNCE-trained checkpoint and embeddings, changing only inference scoring.** InfoNCE + cosine uses the separately InfoNCE-trained checkpoint. Prob abbreviates ProbInfoNCE, whose results use native scoring unless stated otherwise. --- **W1 / Q1(a). How does ProbInfoNCE differ from Zhang et al. [1] and Qiu et al. [2], and why do the differences matter?** **Our novelty concerns the observation model, matching objective, and their connection to CVGL.** Zhang et al. [1] use an input-dependent anchor temperature shared across its candidates. Qiu et al. [2] derive individualized temperatures through KL-distributionally robust optimization, with variables indexed by training examples. We model variation at both image endpoints and cross-view mismatch (Eq. 6), predicting scales from image features for unseen queries and references. Their sum preserves independent-perturbation covariance addition. Appendix C’s same-scale, zero-similarity ratio gives our exact logit (Eq. 32), followed by the specified contrastive NLL (Eqs. 34–37). These differences affect both learning and candidate-dependent retrieval scoring. **Our model, data, and training share this motivation:** editing supplies appearances of the same location, Dual aligns them with one reference, q2q enforces identity consistency, and CAV-Geo evaluates the challenge. A shared temperature-scaling operation does not establish equivalence between these formulations. Tables R1 and R2 test the parameterizations and compare the methods. --- **Q1(b). Do reference-only and alternative combinations separate reference information from the additive form?** As requested, we tested reference-only scaling and alternative combinations on VIGOR Cross-Area. **Addition outperforms every tested alternative on both query conditions (Table R1)**, including multiplication with both image scales. This is because addition is the only one that follows and is supported by Eq. 32 in our submission. **Table R1. Scale parameterizations on VIGOR Cross-Area (R@1, %).** | VIGOR Cross-Area variant | Scoring rule | Original R@1 (%) | FLUX-edited R@1 (%) | | --- | --- | ---: | ---: | | Query-only | One MLP scale per query | 66.27 | 62.10 | | Reference-only | One MLP scale per satellite candidate | 68.51 | 64.62 | | Multiplicative | Product of query and reference scales | 68.67 | 65.07 | | Learned pairwise similarity bias | Learned bias added to the similarity logit | 67.07 | 63.20 | | Additive (ProbInfoNCE, submission Table 7) | Sum of query and reference scales | **70.10** | **66.69** | The one-sided controls use the same MLP architecture on retrieval features. A query-dependent positive scale **changes training but preserves cosine ranks at fixed embeddings**, because it rescales all candidates for that query equally. Reference-only tests candidate information. Multiplication retains both scales and changes their combination. Query-only original is from submission Table 8. Its edited value is a confirmed re-evaluation. The query-only control implements the anchor-scaling idea of Zhang et al. [1]. The reference-only and multiplicative controls show that candidate information alone does not explain the full benefit. --- **Q1(c). How does ProbInfoNCE compare with iSogCLR and other adaptive-temperature methods?** Table R2 directly compares Qiu et al.’s iSogCLR objective [2] and additional alternatives [3, 4, 5, 6, 7], with matched DINOv3 initialization, training data, and optimization settings. **Table R2. Adaptive-temperature methods on VIGOR Cross-Area (R@1, %).** | Method | Mechanism | Original R@1 (%) | FLUX-edited R@1 (%) | | --- | --- | ---: | ---: | | TempNet [3] | Predicts anchor temperatures for a distributionally robust objective | 61.86 | 58.47 | | iSogCLR [2] | Optimizes per-training-example temperatures under that objective | 57.68 | 54.34 | | DySTreSS [4] | Pair-similarity modulation | 29.66 | 26.43 | | MM-TS [5] | Training schedule with a cluster-frequency shift | 48.34 | 42.95 | | TeMo [6] | Pair-similarity modulation with joint cross- and within-view alignment | 62.92 | 48.42 | | AMCL [7], three-head adaptation | Head-wise pair interactions | Training collapse | Training collapse | | ProbInfoNCE (submission Table 7) | Additive image scales | **70.10** | **66.69** |
AuthorsOpenReview ↗

Response to Reviewer aTtE PART2

W2 / Q2. Do the gains hold with another editor, an independent degradation benchmark, or real adverse weather?

Yes—the gains persist on the full test split edited by another generator. Following cross-editor option, we re-edited VIGOR Cross-Area with Qwen-Image-2.1 and evaluated existing checkpoints on the same test locations and retrieval protocol. ProbInfoNCE retains its embedding advantage, and native scoring improves both edited conditions while reducing the drop (Table R3).

Table R3. Cross-editor evaluation and scoring gains on VIGOR Cross-Area (R@1, %).

QueriesInfoNCE + cosineProb + cosineProb + nativeNative gain (pp)Prob drop: cosine / native (pp)
Original65.5769.3370.10+0.77—
FLUX edited61.8064.5666.69+2.134.77 / 3.41
Qwen edited63.0165.4667.36+1.903.87 / 2.74

Both Prob columns use the same checkpoint. Native gain is relative to cosine. Drop is from original to edited queries. Original results are from submission Tables 7 and 9, FLUX combines submission Table 7 with cosine re-evaluation, and Qwen is the new full-test-set evaluation.


W2 / Q3. Can matched InfoNCE Dual and Dual+q2q baselines isolate the loss on original and edited CAV-Geo Cross-Continent queries?

ProbInfoNCE improves both original and edited Cross-Continent retrieval under both matched strategies (Table R4). As requested, InfoNCE includes Dual and Dual+q2q, with matched initialization, original/FLUX training images, optimization, and supervision. Complete Dual+q2q also improves both query conditions on Same-Area.

Table R4. Matched CAV-Geo controls (R@1, %).

CAV-Geo splitTrainingInfoNCE originalInfoNCE editedProbInfoNCE originalProbInfoNCE edited
Cross-ContinentDual14.5811.5819.1115.24
Cross-ContinentDual+q2q16.2012.5120.8217.14
Same-AreaDual61.8457.1965.0657.51
Same-AreaDual+q2q61.6456.6267.3358.91

ProbInfoNCE Dual+q2q values are from submission Table 5. Table R5 is from the VIGOR Cross-Area ablation from submission Table 7, confirming the loss contribution under each training strategy, including original-only training. The effect of q2q depends on the objective, supporting its joint role with ProbInfoNCE in appearance-consistency learning.

Table R5. Matched training strategies on VIGOR Cross-Area (submission Table 7, R@1, %).

Training settingInfoNCE originalInfoNCE editedProbInfoNCE originalProbInfoNCE editedGain, original / edited (pp)
Original only64.7412.2168.0713.82+3.33 / +1.61
MixBatch65.2661.5068.4863.01+3.22 / +1.51
Dual66.9062.9169.2664.01+2.36 / +1.10
Dual + q2q65.5761.8070.1066.69+4.53 / +4.89

Q4(a). What is R@1 in each of the five cities with the full reference gallery?

The Cross-Continent improvement holds in all five test cities, for both original and edited queries (Table R6). As requested, city-wise evaluation retains the full 162,104-reference gallery for every query. Both models use matched Dual+q2q training.

Table R6. Five-city CAV-Geo Cross-Continent results with the full reference gallery (R@1, %).

Test cityQueriesInfoNCE originalInfoNCE editedProbInfoNCE originalProbInfoNCE edited
Chicago25,47917.0012.0723.6118.28
New York27,76912.399.8114.9413.02
San Francisco28,21518.2615.4320.6718.75
Seattle23,75115.2311.4721.1216.05
Vancouver21,56718.5213.8424.9720.19

Q4(b). Are results for other transfer directions available?

CAV-Geo trains outside North America and tests within it. Table R6 establishes consistency across cities in that protocol. Additional transfer-direction results are not available for this response.

AuthorsOpenReview ↗

Response to Reviewer aTtE PART3

We would appreciate a prompt reply indicating whether our responses have resolved your concerns or whether any questions remain, so that we can further address any outstanding issues.

References

[1] Zhang et al. Temperature as Uncertainty in Contrastive Learning. NeurIPS Workshop on Self-Supervised Learning, 2021.
[2] Qiu et al. Not All Semantics are Created Equal: Contrastive Self-supervised Learning with Automatic Temperature Individualization. ICML, 2023.
[3] Qiu et al. To Cool or not to Cool? Temperature Network Meets Large Foundation Models via DRO. ICML, 2024.
[4] Manna et al. Dynamically Scaled Temperature in Self-Supervised Contrastive Learning. IEEE Transactions on Artificial Intelligence, 2025.
[5] Sheludzko et al. MM-TS: Multi-Modal Temperature and Margin Schedules for Contrastive Learning with Long-Tail Data. WACV, 2026.
[6] Duka et al. TeMo: Temperature Modulation for Multimodal Contrastive Learning. arXiv:2609.07540, 2026.
[7] Wang et al. Adaptive Multi-head Contrastive Learning. ECCV, 2024.


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

This paper addresses cross-view geo-localization (CVGL) under viewpoint differences and appearance variations caused by weather and daylight changes. The authors introduce a probabilistic formulation to characterize query-side variation, reference-side variation, and cross-view mismatch. Based on this formulation, they propose ProbInfoNCE, which replaces the shared temperature in InfoNCE with a learned pair-dependent temperature. The paper also introduces FLUX-based appearance augmentation and constructs CAV-Geo benchmark with synthetic weather and daylight variations and a cross-continental evaluation protocol.

Critical strengths
  1. The paper addresses an important practical challenge in CVGL by considering weather and daylight variations, which are often overlooked in existing methods.
  2. The paper provides extensive experiments across multiple benchmarks, including controlled ablation studies, and introduces a large-scale dataset with appearance variations and a challenging cross-continental evaluation protocol.
Critical weaknesses
  1. ProbInfoNCE primarily introduces a pair-dependent temperature into InfoNCE, while adaptive temperature scaling has already been explored in prior work, such as AMCL [1]. The specific novelty of the additive formulation and its advantages for CVGL should be better justified.
  2. The substantial improvements under weather and daylight changes appear to be largely driven by FLUX-based data augmentation. Although Table 7 demonstrates additional benefits from ProbInfoNCE, the paper should more clearly distinguish the contributions of the proposed loss and the additional training data.
  3. Both training and evaluation of weather and daylight robustness rely on FLUX-generated images. Improvements on synthetic appearance changes do not necessarily translate to real-world adverse conditions, leaving the practical robustness of the method insufficiently validated.
  4. The controlled comparisons mainly use DINOv3-initialized ConvNeXt-B. While these experiments partially address the fairness of baseline comparisons, the effectiveness of ProbInfoNCE across different backbones and pretraining strategies remains insufficiently demonstrated.

[1] Wang, Lei, et al. "Adaptive Multi-Head Contrastive Learning." ECCV, 2024.

Questions for authors
  1. What are the key methodological advantages of ProbInfoNCE over existing adaptive temperature methods such as AMCL.
  2. How does the proposed uncertainty-aware contrastive loss specifically address the weather and daylight variations emphasized in the paper's motivation?
  3. How well do the improvements observed on FLUX-generated images generalize to real-world weather and daylight variations?
7 comments
AuthorsOpenReview ↗

Response to reviewer hcxH

We thank the reviewer for constructive feedback.

Scoring definitions. Retrieval tables report R@1 (%), and edited means FLUX unless specified. For L2-normalized query and reference embeddings qi,rjq_i,r_j, cosine ranks satellite candidates by qi⊤rjq_i^\top r_j. Native ranks them by ProbInfoNCE’s learned additive score (submission Eq. 11 and Appendix C, Eq. 32):

\qquad \epsilon=10^{-6}.$$ The two positive scales are predicted by the shared softplus MLP from the corresponding **pre-normalization features**. **Prob + cosine and Prob + native use the same ProbInfoNCE-trained checkpoint and embeddings, changing only inference scoring.** InfoNCE + cosine uses the separately InfoNCE-trained checkpoint. Prob abbreviates ProbInfoNCE, whose results use native scoring unless stated otherwise. --- **W1. What is novel relative to AMCL?** **ProbInfoNCE differs from AMCL in its probabilistic model, objective, and retrieval structure.** AMCL [1] uses head-wise pair uncertainty, computed from interactions between transformed feature vectors, within an MLE-derived objective. Our CVGL analysis separates same-location variation at both endpoints from cross-view mismatch (Eq. 6). Independently predicted image scales retain covariance addition. Their same-scale, zero-similarity density ratio gives our exact logit (Appendix C, Eq. 32) and the specified contrastive NLL (Eqs. 34–37). Retrieval needs only one extra cached scalar per image. **The same motivation connects the model, data, and training:** editing supplies appearances of the same location, Dual aligns them to the same satellite reference, q2q enforces consistency, and CAV-Geo evaluates that challenge. Judging novelty solely by the use of a temperature overlooks this connection and the specific differences from AMCL. Tables R1 and R2 compare the methods and parameterizations. --- **Q1(a). How was AMCL adapted, and how does it compare under matched CVGL settings?** To test the methodological advantage, we adapted AMCL [1] with a DINOv3-initialized CVGL encoder with their implementation. The loss uses the corresponding temperatures for positive and top-k negative similarities, with temperature regularization. Table R1 compares this adaptation and methods [2, 3, 4, 5, 6] on VIGOR Cross-Area under matched initialization, data, and optimization. **Table R1. Adaptive-temperature methods and AMCL on VIGOR Cross-Area (R@1, %).** | Method | Mechanism / training outcome | Original R@1 (%) | FLUX-edited R@1 (%) | | --- | --- | ---: | ---: | | TempNet [2] | Predicts anchor temperatures for a distributionally robust objective | 61.86 | 58.47 | | iSogCLR [3] | Optimizes per-training-example temperatures under that objective | 57.68 | 54.34 | | DySTreSS [4] | Pair-similarity modulation | 29.66 | 26.43 | | MM-TS [5] | Training schedule with a cluster-frequency shift | 48.34 | 42.95 | | TeMo [6] | Pair-similarity modulation with joint cross- and within-view alignment | 62.92 | 48.42 | | AMCL [1], three-head adaptation | Head-wise pair interactions | Training collapse | Training collapse | | ProbInfoNCE (submission Table 7) | Additive image scales | **70.10** | **66.69** | “Training collapse” means the temperature goes to 0 and cosine similarity between any two images is close to 1. --- **Q1(b). What advantage does the additive formulation offer over alternative parameterizations?** **All tested alternative parameterizations also underperform addition on both query conditions (Table R2)**, including multiplication that retains both image scales. **Table R2. Scale parameterizations on VIGOR Cross-Area (R@1, %).** | VIGOR Cross-Area variant | Scoring rule | Original R@1 (%) | FLUX-edited R@1 (%) | | --- | --- | ---: | ---: | | Query-only | One positive MLP scale per query | 66.27 | 62.10 | | Reference-only | One MLP scale per satellite candidate | 68.51 | 64.62 | | Multiplicative | Product of query and reference scales | 68.67 | 65.07 | | Learned pairwise similarity bias | Learned bias added to the similarity logit | 67.07 | 63.20 | | Additive (ProbInfoNCE, submission Table 7) | Sum of query and reference scales | **70.10** | **66.69** | The one-sided controls use the same MLP architecture on retrieval features. A query-dependent positive scale **changes training but preserves cosine ranks at fixed embeddings**, because it rescales all candidates for that query equally. Reference-only tests candidate information. Multiplication retains both scales and changes their combination. Query-only original is from submission Table 8. Its edited value is a confirmed re-evaluation.
AuthorsOpenReview ↗

Response to reviewer hcxH PART2

Q2. How does the loss address weather and daylight variation?

Appearance supervision and ProbInfoNCE address the same location-preserving objective together. Editing varies appearance, Dual aligns both panoramas to the same reference, and q2q enforces their identity consistency. ProbInfoNCE adjusts matching evidence using both images’ scales throughout these objectives. Tables R3 and R4 evaluate their joint role, and Table R5 isolates the inference-scoring benefit under appearance changes.


W2. How much of the gain comes from the loss versus FLUX augmentation?

Table R3 is from Table 7 of our submission and holds training exposure to FLUX-edited images same within each row. ProbInfoNCE improves original and edited retrieval under every strategy, including original-only training. Adding q2q benefits ProbInfoNCE but reduces InfoNCE performance on VIGOR, supporting their joint role in learning appearance consistency.

Table R3. Matched training strategies on VIGOR Cross-Area (submission Table 7, R@1, %).

Training settingInfoNCE originalInfoNCE editedProbInfoNCE originalProbInfoNCE editedGain, original / edited (pp)
Original only64.7412.2168.0713.82+3.33 / +1.61
MixBatch65.2661.5068.4863.01+3.22 / +1.51
Dual66.9062.9169.2664.01+2.36 / +1.10
Dual + q2q65.5761.8070.1066.69+4.53 / +4.89

Matched CAV-Geo controls in Table R4 (InfoNCE retrained with DINOv3 and augmented data) show gains under both Cross-Continent strategies. Complete ProbInfoNCE with Dual+q2q improves both conditions on both splits.

Table R4. Matched CAV-Geo controls (R@1, %).

CAV-Geo splitTrainingInfoNCE originalInfoNCE editedProbInfoNCE originalProbInfoNCE edited
Cross-ContinentDual14.5811.5819.1115.24
Cross-ContinentDual+q2q16.2012.5120.8217.14
Same-AreaDual61.8457.1965.0657.51
Same-AreaDual+q2q61.6456.6267.3358.91

ProbInfoNCE Dual+q2q values are from submission Table 5. Comparisons match initialization, training images, optimization settings, and supervision within each strategy.


W3 / Q3. Do the gains on FLUX edits generalize to real weather and daylight changes?

The gains generalize beyond FLUX to a different image editor. We re-edited the full VIGOR Cross-Area test split with Qwen-Image-2.1 and evaluated existing checkpoints. The embedding advantage persists, and native scoring further improves retrieval while reducing the edited-query drop (Table R5).

Table R5. Cross-editor evaluation and scoring gains on VIGOR Cross-Area (R@1, %).

QueriesInfoNCE + cosineProb + cosineProb + nativeNative gain (pp)Prob drop: cosine / native (pp)
Original65.5769.3370.10+0.77—
FLUX edited61.8064.5666.69+2.134.77 / 3.41
Qwen edited63.0165.4667.36+1.903.87 / 2.74

Both Prob columns use the same checkpoint. Native gain is relative to cosine. Drop is from original to edited queries. Original results are from submission Tables 7 and 9, FLUX combines submission Table 7 with cosine re-evaluation, and Qwen is the new full-test-set evaluation.

Large-scale, geographically aligned panoramas at the same GPS locations across diverse real weather and daylight conditions require repeated collection and verification. The substantial time, funding, and personnel costs motivate controlled editing. Qwen tests cross-editor transfer. Transfer to naturally captured adverse-weather panoramas remains unmeasured here.


W4. Do the gains persist across backbones and pretraining strategies?

The gains persist with ImageNet-1K/22K-initialized backbones. To address your backbone/pretraining concern, we evaluated ConvNeXt-B with ImageNet-1k/22K initialization (Sample4Geo uses 22K checkpoint) and ResNet-50 with ImageNet-1K initialization on VIGOR Cross-Area. Table R6 compares objectives within each matched training configuration, with the DINOv3 row from submission Table 7.

Table R6. Backbones and initialization on VIGOR Cross-Area (R@1, %).

Backbone / initializationInfoNCE originalInfoNCE editedProbInfoNCE originalProbInfoNCE edited
ConvNeXt-B / DINOv3 (submission Table 7, Dual+q2q)65.5761.8070.1066.69
ConvNeXt-B / ImageNet-1K58.9953.9559.9255.20
ConvNeXt-B / ImageNet-22K63.4158.1666.4761.20
ResNet-50 / ImageNet-1K25.9921.3032.9026.67
AuthorsOpenReview ↗

Response to reviewer hcxH PART3

We would appreciate a prompt reply indicating whether our responses have resolved your concerns or whether any questions remain, so that we can further address any outstanding issues.

References

[1] Wang et al. Adaptive Multi-head Contrastive Learning. ECCV, 2024.
[2] Qiu et al. To Cool or not to Cool? Temperature Network Meets Large Foundation Models via DRO. ICML, 2024.
[3] Qiu et al. Not All Semantics are Created Equal: Contrastive Self-supervised Learning with Automatic Temperature Individualization. ICML, 2023.
[4] Manna et al. Dynamically Scaled Temperature in Self-Supervised Contrastive Learning. IEEE Transactions on Artificial Intelligence, 2025.
[5] Sheludzko et al. MM-TS: Multi-Modal Temperature and Margin Schedules for Contrastive Learning with Long-Tail Data. WACV, 2026.
[6] Duka et al. TeMo: Temperature Modulation for Multimodal Contrastive Learning. arXiv:2609.07540, 2026.

Reviewer_qcXDOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors address the cross-view geo-localization problem by arguing that using a fixed temperature for all image pairs in contrastive learning is suboptimal. Their motivation comes from a probabilistic decomposition of the expected squared embedding distance between matching street-view and satellite images into query-side variation, reference-side variation, and cross-view mismatch. The authors argue that previous research has primarily focused on reducing cross-view mismatch, while paying less attention to the other sources of variation. To address this gap, they propose a learned, pair-dependent temperature for contrastive learning. Additionally, they introduce a new dataset by combining the existing VIGOR and DRESS datasets and synthetically generating appearance variations. Their proposed method achieves competitive or superior performance compared to state-of-the-art approaches across established CVGL benchmarks.

Critical strengths

Strong empirical performance

The proposed approach yields consistent gains across the VIGOR, CVUSA, and CVACT benchmarks. The authors also conduct extensive ablation studies and controlled baseline evaluations to show the effectiveness of their contributions.

The new dataset

Handling weather and illumination variations remains an important challenge in CVGL. These variations are underrepresented in most popular CVGL benchmarks, although some recent works have also introduced more realistic benchmarks with appearance variations [1]. Therefore, extending existing datasets to include these variations is a valuable contribution.

Computational efficiency

The proposed method only adds a lightweight MLP to a straightforward shared-encoder retrieval architecture, introducing negligible computational overhead.

Additional ablations in the appendix

The authors provide further experiments in the appendix, including comparisons against an existing input-dependent temperature approach, demonstrating the advantages of their proposed formulation.

[1] Erzurumlu et al. (2026), Just Zoom In: Cross-View Geo-Localization via Autoregressive Zooming.

Critical weaknesses

Limited methodological novelty

Input-dependent temperature scaling is not a novel concept and has been examined in multiple prior works [2–7]. Specifically, Qiu et al. [3, 4] explored individualized temperatures and trained temperature prediction networks for general contrastive learning. Other recent techniques have similarly investigated adaptive temperature scaling grounded in semantic density [6] and pairwise similarity [7]. The primary methodological distinction of this paper lies in the additive combination of query-side and reference-side temperatures. However, this appears to be an incremental extension of established adaptive temperature approaches rather than a major methodological contribution.

While the authors motivate their additive formulation using the covariance addition property of independent Gaussian perturbations, it remains unclear why this specific design is preferable to existing methods. Although the authors compare against [2], it would be valuable to include controlled comparisons with established adaptive temperature techniques [3, 4] as well as more recent approaches [5, 7]. Furthermore, comparisons against simpler alternative parameterizations, such as multiplicative temperature scaling or learned similarity biases, would help clarify whether the performance gains stem specifically from the proposed additive uncertainty formulation or simply from adding flexibility to the similarity function.

Even if the proposed formulation outperforms these alternatives, its methodological novelty remains somewhat limited given the existing literature on adaptive temperature scaling.

Weak connection between probabilistic motivation and the proposed method

The authors present an interesting probabilistic motivation. However, its connection to the proposed method is not sufficiently established. The appendix provides a Gaussian likelihood-ratio motivation for the proposed logit, but several additional assumptions and modeling choices are introduced along the way, including isotropic uncertainty and a zero-similarity reference distribution. These choices do not necessarily follow from the initial bias–variance decomposition.

Furthermore, the learned temperature scores are neither direct estimates of covariance traces nor constrained to represent calibrated uncertainty. This leaves it unclear whether the method is truly uncertainty-aware or simply an empirically effective adaptive scoring mechanism.

Unclear source of improvements from ProbInfoNCE

Although ProbInfoNCE consistently improves retrieval performance over standard InfoNCE, the source of these improvements remains unclear. In particular, the authors motivate their method by arguing that adaptive temperature scaling helps handle sample-dependent ambiguities caused by factors such as weather and illumination. However, the experiments do not convincingly establish that the observed improvements arise specifically from better handling of these ambiguities.

The improvements could instead result from more flexible contrastive optimization or changes to the learned similarity function, independently of appearance variation.

Analyzing how predicted temperature scores and retrieval performance shift across varying levels of appearance variation would help confirm if the mechanism truly resolves its stated motivation.

Limitations of synthetic data generation and evaluation

CAV-Geo is a valuable contribution. However, its appearance variations are generated using FLUX, which introduces two important limitations.

First, generative editing may alter geographically discriminative details. The proposed similarity-based checks cannot guarantee geographical or geometric consistency.

Second, synthetically generated images may follow a different distribution from real-world images, beyond the intended weather and illumination changes. Therefore, the substantial performance degradation of existing methods on the synthetic benchmark may partly reflect a domain shift introduced by the generative model itself. Models trained on real images could perform poorly on these synthetic images even if they are reasonably robust to genuine weather and illumination variations.

For instance, Sample4Geo’s R@1 drops from 57.64% on original queries to 3.00% on synthetic queries, whereas the proposed method achieves 58.91% on edited queries. However, because the proposed model is explicitly trained on FLUX-generated images and the baseline is not, this gap may reflect adaptation to the synthetic distribution rather than superior handling of real-world appearance variations.

It would be valuable to evaluate the methods on real-world appearance variations, even using a relatively small dataset, or on synthetic images generated by a different model. Such experiments would help determine whether the reported robustness improvements generalize beyond the specific image-generation pipeline.

[2] Zhang et al. (2021), Temperature as Uncertainty in Contrastive Learning. [ 3] Qiu et al. (2023), Not All Semantics are Created Equal: Contrastive Self-supervised Learning with Automatic Temperature Individualization.

[4] Qiu et al. (2024), To Cool or not to Cool? Temperature Network Meets Large Foundation Models via DRO.

[5] Manna et al. (2025), Dynamically Scaled Temperature in Self-Supervised Contrastive Learning.

[6] Sheludzko et al. (2026), MM-TS: Multi-Modal Temperature and Margin Schedules for Contrastive Learning with Long-Tail Data.

[7] Duka et al. (2026), TeMo: Temperature Modulation for Multimodal Contrastive Learning.

Questions for authors
  1. Why is the proposed additive uncertainty parameterization preferable to existing adaptive temperature methods [3–7]? Could the authors provide controlled comparisons with these approaches, particularly the established methods [3, 4], or simpler alternative parameterizations such as multiplicative temperature scaling?

  2. What does the probabilistic interpretation provide beyond motivating the proposed formulation? Does the probabilistic model lead to any empirically verifiable predictions? How does this interpretation relate to simply learning an MLP for adaptive similarity scaling?

  3. What do the learned temperature scores actually reflect? Do they capture ambiguities caused by appearance variations, geographical ambiguity, or both? For example, do locations with repetitive layouts systematically exhibit different predicted scales compared to geographically distinctive locations?

  4. What is the contribution of temperature adaptation during training versus inference? Appendix Table 9 already shows that pair-dependent inference improves R@1 across all nine reported settings. A more detailed decomposition would help determine how much of the overall improvement comes from learning better representations versus modifying the similarity scores at inference.

8 comments
AuthorsOpenReview ↗

Response to qcXD

We thank the reviewer for constructive feedback.

Scoring definitions. Retrieval tables report R@1 (%), and edited means FLUX unless specified. For L2-normalized query and reference embeddings qi,rjq_i,r_j, cosine ranks satellite candidates by qi⊤rjq_i^\top r_j. Native ranks them by ProbInfoNCE’s learned additive score (submission Eq. 11 and Appendix C, Eq. 32):

\qquad \epsilon=10^{-6}.$$ The two positive scales are predicted by the shared softplus MLP from the corresponding **pre-normalization features**. **Prob + cosine and Prob + native use the same ProbInfoNCE-trained checkpoint and embeddings, changing only inference scoring.** InfoNCE + cosine uses the separately InfoNCE-trained checkpoint. Prob abbreviates ProbInfoNCE, whose results use native scoring unless stated otherwise. --- **W1. What is novel beyond existing adaptive-temperature methods?** **Our novelty is the CVGL-specific probabilistic analysis and its endpoint-additive matching model.** Eq. 6 identifies variation at both endpoints and their mean mismatch. Under Appendix C’s independent isotropic perturbation model, the same-scale, zero-similarity density ratio gives our exact logit (Eq. 32), followed by the contrastive NLL (Eqs. 34–37). The same motivation organizes appearance editing, Dual alignment, q2q consistency, and CAV-Geo around location identity. **Reducing this unified contribution to the presence of a temperature overlooks the central modeling connection.** The cited methods differ in their variables and objectives. TempNet and iSogCLR [1, 2] derive anchor temperatures from KL-distributionally robust optimization, using a predictor or training-example variables. DySTreSS [3] modulates cosine-based temperature to control contrastive repulsion. MM-TS [4] uses a cluster-frequency-dependent training schedule. TeMo [5] uses cosine-based modulation with cross- and within-modal alignment. AMCL [6] models head-wise pair uncertainty through feature interactions and an MLE objective. We predict each observation’s scale separately and combine them according to the perturbation model. Equal cosine can yield different scores. **Shared temperature scaling does not make these models, objectives, or motivations equivalent.** The addition is not a random design but well-supported by our probabilistic analysis (see W2 / Q2(a–b)) and Table R2. --- **Q1(a). How does ProbInfoNCE compare with existing adaptive-temperature methods under matched settings?** As requested, we compared methods [1, 2, 3, 4, 5], multiplicative scaling, and learned pairwise bias on VIGOR Cross-Area, and additionally evaluated AMCL [6], with matched DINOv3 initialization, training data, and optimization settings. Table R1 reports the prior-method comparison, and Table R2 tests the scoring structure. AMCL collapes to near zero temperature and near 1 similarity from any two inputs. **Table R1. Adaptive-temperature methods on VIGOR Cross-Area (R@1, %).** | Method | Mechanism | Original R@1 (%) | FLUX-edited R@1 (%) | | --- | --- | ---: | ---: | | TempNet [1] | Predicts anchor temperatures for a distributionally robust objective | 61.86 | 58.47 | | iSogCLR [2] | Optimizes per-training-example temperatures under that objective | 57.68 | 54.34 | | DySTreSS [3] | Pair-similarity modulation | 29.66 | 26.43 | | MM-TS [4] | Training schedule with a cluster-frequency shift | 48.34 | 42.95 | | TeMo [5] | Pair-similarity modulation with joint cross- and within-view alignment | 62.92 | 48.42 | | AMCL [6]| Head-wise pair interactions | Training collapse | Training collapse | | ProbInfoNCE (submission Table 7) | Additive image scales | **70.10** | **66.69** |
AuthorsOpenReview ↗

Response to qcXD PART 2

Q1(b). Why use addition rather than multiplication or a learned similarity bias?

Every tested alternative parameterization underperforms addition on both query conditions (Table R2). This is because only addition follows our formulation, indicating that the addition is not a random design but well-supported by our probabilistic analysis.

Table R2. Scale parameterizations on VIGOR Cross-Area (R@1, %).

VIGOR Cross-Area variantScoring ruleOriginal R@1 (%)FLUX-edited R@1 (%)
Query-onlyOne positive MLP scale per query66.2762.10
Reference-onlyOne MLP scale per satellite candidate68.5164.62
MultiplicativeProduct of query and reference scales68.6765.07
Learned pairwise similarity biasLearned bias added to the similarity logit67.0763.20
Additive (ProbInfoNCE, submission Table 7)Sum of query and reference scales70.1066.69

The one-sided controls use the same MLP architecture on retrieval features. A query-dependent positive scale changes training but preserves cosine ranks at fixed embeddings, because it rescales all candidates for that query equally. Reference-only tests candidate information. Multiplication retains both scales and changes their combination. Query-only original is from submission Table 8 and edited is re-evaluated.


W2 / Q2(a). What does the probabilistic model provide beyond an adaptive MLP?

The model specifies a score with an exact likelihood-ratio interpretation, beyond an unconstrained adaptive MLP. For unit retrieval embeddings, let cij=qi⊤rjc_{ij}=q_i^\top r_j. Following Eq. 11,

\qquad \epsilon=10^{-6}.$$ Here $\sigma_\theta(q_i)$ and $\sigma_\theta(r_j)$ abbreviate predictions from the corresponding **pre-normalization features** using the shared softplus MLP. They are variance-like effective scales, not Gaussian standard deviations. Under its independent isotropic perturbation model, Appendix C takes $\Delta_{ij}\sim\mathcal N(q_i-r_j,\tau_{ij}I_d)$. Since $\|q_i-r_j\|^2=2-2c_{ij}$, its log-density at zero, for a pair-independent constant $C$, is $C+c_{ij}/\tau_{ij}-1/\tau_{ij}-(d/2)\log\tau_{ij}$. Taking the ratio against the same-scale, zero-cosine reference gives exactly Eq. 32: $$\log\frac{p(\Delta_{ij}=0\mid q_i,r_j)} {p_0(\Delta_{ij}=0\mid\tau_{ij})} =\frac{c_{ij}}{\tau_{ij}}=g_{ij}.$$ --- **W2 / Q2(b). What do the assumptions imply about uncertainty, and what can be tested?** Here $p$ and $p_0$ denote **density values**, not probability masses. The ratio cancels scale-only terms that would not cancel in an ordinary candidate softmax. Isotropy enables one scalar per image. Zero cosine defines the reference, not a claim that real negatives are orthogonal. **These assumptions specify the derivation’s scope without weakening the exact equality within that model.** Eqs. 34–37 give the specified softmax matching model’s bidirectional NLL. Eq. 13 adds scale regularization. The learned scales need not be calibrated covariance estimates for this correspondence to hold. The multiplicative and additive rows of Table R2 provide the direct structural test, retaining both image scales while changing their combination. **Only addition among these controls implements Eq. 32, and it performs better under both query conditions.** --- **Q3. Do the scales reflect appearance variation, geographic ambiguity, or both?** **The scales reflect retrieval-relevant variation, which can include appearance changes and geographic ambiguity.** They are learned through location matching. Table R3 is from Appendix G’s association between larger mean query scales and unsuccessful retrievals. **Table R3. Query scale and retrieval outcome on VIGOR Cross-Area (Appendix G).** | VIGOR Cross-Area retrieval outcome (Appendix G) | Mean query scale | | --- | ---: | | Correct top-1 retrieval | 0.039 | | Incorrect top-1 retrieval | **0.043** | ---
AuthorsOpenReview ↗

Response to qcXD PART 3

W3. How does adaptive scoring help under appearance changes?

At fixed ProbInfoNCE weights, native scoring improves edited retrieval and reduces the original-to-edited drop under both editors compared with cosine scoring (Table R4, scoring-gain and drop columns). This is a inference time only comparison.

Table R4. Embedding and scoring gains across query conditions on VIGOR Cross-Area (R@1, %).

QueriesInfoNCE + cosineProb + cosineProb + nativeEmbedding gain (pp)Scoring gain (pp)Prob R1 drop on edited: cosine / native (pp)
Original65.5769.3370.10+3.76+0.77—
FLUX edited61.8064.5666.69+2.76+2.134.77 / 3.41
Qwen edited63.0165.4667.36+2.45+1.903.87 / 2.74

Embedding gain compares the two cosine columns, which shows a better raw embedding space on ProbInfoNCE (simple cosine similariy still outperforms InfoNCE). Scoring gain compares native and cosine on the same ProbInfoNCE checkpoint. Drop is from original to edited queries. Original results are from submission Tables 7 and 9. FLUX combines submission Table 7 with cosine re-evaluation, and Qwen uses the full re-edited test split.

These results support useful score adjustment under appearance change.


Q4. How much of the R@1 gain comes from better embeddings versus adaptive scoring at inference?

Both learned embeddings and inference scoring contribute in every tested condition (Table R4, embedding-gain and scoring-gain columns). As requested, the comparison matches DINOv3 initialization, training data, supervision, and optimization. Cosine-to-cosine compares only the raw embedding spaces trained via InfoNCE and ProbInfoNCE. Switching the same ProbInfoNCE checkpoint to native scoring isolates the additional inference-time temperature scaling gain. The scoring benefit is larger on edited queries, indicating the learned scale does benefit the appearance variation.


W4(a). Do the gains generalize beyond the FLUX editor?

The gains persist with a different test editor. Following the cross-editor option, we re-edited the full VIGOR Cross-Area test split with Qwen-Image-2.1 and evaluated the existing checkpoints (Table R4, Qwen-edited row).

Collecting geographically aligned, same-location panoramas across real weather conditions requires substantial time, funding, and personnel. This constraint motivates editing. Qwen establishes cross-editor generalization. Naturally captured adverse-weather transfer remains unmeasured here.


W4(b). Do gains persist with equal FLUX training exposure?

Table R5 is from submission Table 7, matching training images and alignment strategy within each row. ProbInfoNCE improves both conditions under every strategy, including original-only training, establishing an objective contribution beyond additional FLUX exposure.

Table R5. Matched training strategies on VIGOR Cross-Area (submission Table 7, R@1, %).

Training settingInfoNCE originalInfoNCE editedProbInfoNCE originalProbInfoNCE editedGain, original / edited (pp)
Original only64.7412.2168.0713.82+3.33 / +1.61
MixBatch65.2661.5068.4863.01+3.22 / +1.51
Dual66.9062.9169.2664.01+2.36 / +1.10
Dual + q2q65.5761.8070.1066.69+4.53 / +4.89

The CAV-Geo controls in Table R6 confirm gains under both Cross-Continent strategies (InfoNCE is trained with same DINOv3 initialization, training images, optimization settings.). The complete Dual+q2q method improves both conditions on both splits.

Table R6. Matched CAV-Geo controls (R@1, %).

CAV-Geo splitTrainingInfoNCE originalInfoNCE editedProbInfoNCE originalProbInfoNCE edited
Cross-ContinentDual14.5811.5819.1115.24
Cross-ContinentDual+q2q16.2012.5120.8217.14
Same-AreaDual61.8457.1965.0657.51
Same-AreaDual+q2q61.6456.6267.3358.91

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response

Our work starts from one question: how can cross-view geo-localization preserve location identity despite viewpoint and appearance variation? Our probabilistic analysis decomposes same-location feature mismatch into variation at both image endpoints and their mean mismatch (Eq. 6). This motivates modeling both endpoints in ProbInfoNCE. Under Appendix C’s stated perturbation model and same-scale reference, the additive score is exactly the log-density ratio in Eq. 32, and Eqs. 34–37 give the specified matching model’s contrastive NLL.

The same analysis guides the supervision and data design. Appearance editing supplies different observations of a fixed location. Dual Matrix aligns them with the same satellite reference. The q2q objective enforces consistency of their location identity. These objectives work with ProbInfoNCE to learn matching under variation. CAV-Geo evaluates this shared challenge at scale and across continents. Controlled editing addresses the substantial time, funding, and personnel costs of collecting geographically aligned panoramas at the same locations across diverse weather and daylight conditions. The probabilistic model, training strategy, and benchmark are complementary parts of one approach, grounded in the same motivation and analysis. Their contribution lies in this connection, and the ablations test their roles and interactions within it.

To address the reviews, we added comparisons with prior adaptive-temperature methods and alternative scale parameterizations, matched CAV-Geo controls and five-city results, and additional backbone/initialization comparisons. We also re-edited the full VIGOR Cross-Area test split with Qwen-Image-2.1 and separated embedding gains from native-scoring gains using the same checkpoints. These experiments test the modeling choices, training interactions, and generalization of the unified approach.

We will respond to reviewers' questions individually.

Dear reviewers,

We have updated the rebuttal revision with an appendix section N. We kindly ask for further comments and discussions if there are concerns need to be addressed. We appreciate all reviewers patience and constructive feedbacks.

AuthorsOpenReview ↗

Summary of the Rebuttal Period

We appreciate reviewers' patience and constructive feedbacks. We briefly summarize the discussions below

  1. Reviewers highlighted ProbInfoNCE’s lightweight pair-dependent scoring and consistent retrieval gains across CVGL benchmarks. They also recognized CAV-Geo’s large-scale weather and day–night evaluation and challenging cross-continent protocol as valuable contributions. hcxH specifically recognized the value of generative image editing for appearance augmentation.

  2. The main concerns were how ProbInfoNCE compares with existing adaptive-temperature methods, whether the additive form drives the gains, and whether the gains generalize across editors, backbones, and test cities. Reviewers also asked us to separate the effects of the loss, data augmentation, and inference-time scoring.

  3. We addressed the main concerns raised by qcXD and hcxH, as well as aTtE’s experimental requests, with matched adaptive-temperature comparisons, scale-combination ablations, full-split Qwen evaluation (cross-editor generalization), training-versus-inference analysis, matched CAV-Geo baselines, backbone and initialization comparisons, and full-gallery results for all five CAV-Geo test cities (Appendix N, Tables 15–20). Both qcXD and hcxH raised their scores from 2 to 3. aTtE acknowledged that the new results further support the method’s effectiveness and maintained a concern about novelty.

  4. Reviewer aTtE pointed to prior work on adaptive temperature and appearance augmentation. Prior adaptive-temperature methods use anchor-level temperatures or learn pair temperatures from feature interactions. Our two-image perturbation model derives the exact additive logit (Eq. 32) and the contrastive NLL (Eqs. 34–37) for the specified matching model. ProbInfoNCE achieves +7.18 pp in R@1 over TeMo on original VIGOR Cross-Area queries, outperforms every tested alternative scale combination, and achieves state-of-the-art R@1 in our VIGOR Cross-Area benchmark comparison (Tables 2 and 15–16). These results support the probabilistic motivation and the effectiveness of the mathematically derived additive form.

  5. One probabilistic motivation unifies the algorithmic and dataset contributions. Both aim to preserve location identity under viewpoint and appearance variation. The algorithmic contribution models variation on both sides of a match and uses location-consistent supervision across original and edited views. The dataset contribution, CAV-Geo, makes the same problem measurable through controlled appearance changes and cross-continent retrieval.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.