acceptodds
Under review as a conference paper at ICLR 2027

GeoSync: Unified Representation Learning for End-to-End Cross-View Geolocalization

Abstract

We propose GeoSync, a unified framework for cross-view image geolocalization that jointly learns satellite image retrieval and fine-grained localization from a shared cross-view representation. Existing approaches typically separate these tasks, using global image-level matching to retrieve a satellite reference and a dedicated localizer to estimate the position within it. This decoupling limits interaction between retrieval and localization while adding computation during inference and increasing system cost. GeoSync learns a shared representation that captures global scene understanding for retrieval and local feature correlations for precise localization. We further introduce consistency supervision across positive and semi-positive satellite references and a sub-meter refinement module to improve accuracy and robustness. Localization-aware re-ranking then uses localization evidence to refine retrieved candidates, allowing localization to reinforce retrieval. Experiments on VIGOR and DReSS demonstrate strong performance in both retrieval and fine-grained localization, yielding accurate and fast end-to-end geolocalization from a single shared representation and maintaining robustness as the target moves toward the boundary of the satellite reference.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Cross-View Image Geolocalization, Representation Learning, Metric Learning, Fine-Grained Localization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_W7daOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoSync, a unified and simplified cross-view retrieval–localization framework. Ground panoramas and satellite images are encoded separately, and the retrieval and localization tasks mainly reuse the encoder outputs: MCG estimates a coarse location via multi-scale correlation, SMRR refines the local position, and LAR re-ranks retrieval candidates using feature consistency at the predicted location, ultimately outputting geographic coordinates. The problem studied is practically valuable: how to perform accurate end-to-end localization at low online cost when the retrieved satellite reference may be off-center relative to the query location. However, the current manuscript still insufficiently clarifies the novelty boundary of the method, the sources of its performance gains, and the evidence for bidirectional task reciprocity.

Critical strengths
  • The framework is compact and has engineering value. Retrieval, coarse localization, and refinement mainly reuse encoded features, avoiding a full independent model for each task. LAR uses fixed candidate verification and score fusion rules without an additional trained re-ranker. This design helps simplify inference.

  • The evaluation covers the practical retrieval–localization pipeline. The paper reports not only localization error given the correct reference but also final geographic error after retrieval. Under the reported settings, GeoSync shows good end-to-end performance; with precomputed gallery features, its online latency is also competitive.

  • The focus on off-center references is practically meaningful. Real retrieval does not guarantee that the query lies at the center of the satellite tile. The paper addresses this issue through semi-positive reference supervision and cross-tile constraints, which is closer to practical use than evaluating localization accuracy only on the main positive reference.

Critical weaknesses

W1. MCG has a direct mechanistic correspondence with AuxGeo’s PCM, and the manuscript does not sufficiently clarify the novelty boundary

AuxGeo’s PCM and its official implementation already include the following computation: constructing a three-level average pooling pyramid from satellite local features → cosine matching between each scale and the ground/BEV global descriptor → low-temperature softmax → coordinate expectation → Euclidean distance supervision on positions for two types of queries. This closely corresponds to the core computational structure of MCG in Section 3.2. The currently identifiable changes include the backbone, the readout of global descriptors, feature resolution, temperature, and scale weights. However, the manuscript does not explain what new computational mechanism or learning constraint MCG introduces beyond these changes.

To be fair, AuxGeo uses this type of positional constraint as auxiliary supervision during retrieval training, whereas GeoSync further uses a similar readout for inference-time localization and connects it to refinement and candidate verification. This is a meaningful system extension, but it does not automatically make MCG itself a new localization mechanism.

Therefore, my main concern is not whether the authors cite AuxGeo, but that the manuscript does not sufficiently explain the direct relationship between MCG and PCM, nor clearly distinguish inherited mechanisms, implementation changes, and new contributions.

W2. The specific increment of LAR has not been isolated and validated

Using fine-grained matching evidence to re-rank global retrieval candidates has precedent. For example, UnifyGeo’s public inference pipeline already includes candidate retrieval, fine-grained matching re-ranking, and final localization. GeoSync’s LAR uses feature similarity at the refined location, together with within-candidate normalization and fixed fusion weights. This design is reasonable, but the current experiments mainly show that adding LAR is effective; they do not sufficiently isolate whether the gain comes from:

  • verification at the refined location, rather than simpler local scores such as correlation-map peaks;
  • within-candidate normalization or fusion weights;
  • features better suited to verification;
  • a larger candidate set.

Without a comparison of scoring rules under fixed features and a fixed candidate set, LAR’s independent value relative to existing local-verification re-ranking remains unclear.

W3. Cross-method comparisons support system-level performance but are insufficient to isolate the advantage of shared representations

Appendix B.2 states that the cascade baselines use only the top-1 reference, UnifyGeo re-ranks 5 candidates by default, and GeoSync uses 10 or 20 candidates. Different candidate numbers are not inherently unfair. If GeoSync can process more candidates at lower online cost, that is a legitimate system advantage. However, such a comparison cannot by itself prove that direct shared readout is better than task-specific branches, because the candidate budget also changes.

The current experiments need to distinguish two questions more clearly: whether GeoSync is a better system, and whether the shared structure is the main cause of its performance gain. At minimum, a targeted structural comparison under the same pretrained backbone, same supervision, and same training recipe is needed, rather than requiring retraining of all existing methods.

Questions for authors
  • What are the substantive changes of MCG relative to PCM?

Beyond the backbone, descriptor readout, scale, and hyperparameters, is there a new computational or supervision mechanism? How much of GeoSync’s localization gain can be explained by directly using a PCM-style readout at inference?

  • Where does LAR’s key gain come from?

Under fixed features and a fixed candidate set, does similarity at the refined location outperform simpler scores such as the maximum local correlation response? How much do normalization and fusion weights contribute separately?

  • Has the shared structure been validated under the same backbone?

Under the same pretraining, supervision, and training recipe, what are the differences in accuracy and resource cost between direct shared readout and ``shared backbone + task-specific branches”?

4 comments
AuthorsOpenReview ↗

Response to Reviewer W7da (1/2)

We thank the reviewer for the detailed feedback. We answer each point with a controlled experiment on VIGOR cross-area.

W1 / Q1. MCG and AuxGeo's PCM

MCG uses the same kind of correlation readout as PCM, and we now state this in Sec. 3.2. But MCG is not a PCM readout moved to inference, and the readout is not what we claim as new. Our novelty is the joint representation: one descriptor is learned for both retrieval and fine-grained localization, and MCG is how fine-grained localization reads it. The differences from PCM are in its role, its supervision and its use:

  • Role. In AuxGeo, PCM is an auxiliary loss for retrieval and is discarded at inference. In GeoSync, MCG is the coarse stage of fine-grained localization, and its query is the retrieval descriptor itself. This is the design choice that couples the two tasks.
  • Supervision. MCG is supervised on both primary and semi-positive references, so it works on any reference that covers the query.
  • Use. SMRR refines the MCG estimate, and LAR uses the refined position to verify the retrieved candidates.

To measure how much a PCM-style readout explains, we used AuxGeo's official model with its PCM directly for fine-grained localization (Table R3).

Table R3: AuxGeo's PCM used for fine-grained localization. Ground-truth satellite reference, Primary and Random reference protocols of Table 7. Best in bold.

MethodRandom reference: Med. ↓Random reference: R@1m ↑Random reference: R@10m ↑Primary reference: Med. ↓Primary reference: R@1m ↑Primary reference: R@10m ↑
AuxGeo with its PCM15.214.5540.352.4414.7791.76
GeoSync, MCG2.0119.6091.861.8022.5794.35
GeoSync, MCG + SMRR1.6625.8193.541.4730.4895.36

PCM alone is insufficient. Used at inference, PCM works only on the primary reference (2.44 m) and fails under the Random protocol (15.21 m). GeoSync's MCG reaches 2.01 m there, and SMRR brings it to 1.66 m.

W2 / Q2. Where the gain of LAR comes from

Mainly from the jointly learned features. We applied LAR, unchanged, to the retrieval-only variant of Table 5. It has the same architecture, so LAR can be computed, but its features were trained with the retrieval loss only (Table R4).

Table R4: The same LAR on features learned with and without joint training. Best in bold.

TrainingHit Rate, no re-rankingHit Rate, with LARChange
Retrieval-only (Table 5)76.6662.96-13.70
Joint (GeoSync)75.1583.60+8.45

Joint learning trains the satellite features to agree with the query at its true location. Agreement at the predicted location therefore becomes evidence that the reference is correct. Retrieval-only features never receive this signal, so the same verification hurts. These features also make the verification score a much better detector of wrong retrievals than the retrieval similarity (AUROC 0.904 vs. 0.793).

The scoring rule adds the rest (Table R5, same features and same K = 20 candidates).

Table R5: Design of the LAR score. Best in bold.

ScoreR@1 ↑Hit Rate ↑
No re-ranking63.3575.15
Verification at the coarse (MCG) position67.5882.94
LAR without normalization66.8582.03
LAR with fusion weight 168.2381.67
LAR (refined position, normalized, weight 2)69.0983.60

Verifying at the refined position, the normalization and the fusion weight each contribute about 1 to 2 points. Other simple local scores, such as the maximum correlation response, give similar gains on these features, which again points to the features and not to the rule. A larger candidate set is not the source either: with K = 5, LAR already gives +7.03 of the +8.45 Hit Rate points.

AuthorsOpenReview ↗

Response to Reviewer W7da (2/2)

W3 / Q3. Candidate budget, and shared readout vs. task-specific branches

Same candidate budget as UnifyGeo. We ran both methods with the same K, using UnifyGeo's released model and code (Table R6).

Table R6: Re-ranking with the same candidate budget. R@1 / Hit Rate. Best in bold.

Candidates KUnifyGeoGeoSync
No re-ranking61.17 / 68.2463.35 / 75.15
5 (UnifyGeo's default)67.58 / 72.2767.90 / 82.18
1067.47 / 72.2468.72 / 83.16
2067.20 / 71.9269.09 / 83.60

With the same K = 5, GeoSync is 9.91 Hit Rate points higher. The two methods also differ in kind. UnifyGeo is a multi-task design: each task has its own heads and decoder, and re-ranking is a learned step that does not improve with more candidates. GeoSync learns one representation jointly, so LAR needs no training, improves with more candidates, and costs 0.08 ms for 20 of them. Verification that scales at no cost matters in practice. It also gives better end-to-end accuracy in the challenging cross-area setting (82.56 vs. 72.65 R@10m, Table 1), which comes from the robust jointly learned representation.

Shared readout vs. task-specific branches. We ran the requested comparison, with the same pretrained DINOv3, supervision and recipe (Table R7). In GeoSync, one descriptor (the CLS token) is both the retrieval descriptor and the query of fine-grained localization. In the task-specific version, two separate MLP heads on this token produce one descriptor per task. In all variants the backbone is trained by both losses, and the only difference is whether the two tasks read the same descriptor.

Table R7: Shared readout vs. task-specific heads. Same backbone, supervision and training recipe. Best in bold.

ReadoutExtra paramsRetrieval: R@1 / HitRetrieval: after LAREnd-to-end Med. ↓End-to-end R@1m ↑End-to-end R@5m ↑End-to-end R@10m ↑
Shared descriptor (GeoSync)none63.35 / 75.1569.09 / 83.601.6327.2480.3882.56
One shared head (control)+0.30M62.60 / 74.3868.85 / 83.561.6427.1580.2682.54
Two task-specific heads+0.59M57.42 / 69.5865.72 / 81.451.8025.9278.4180.73

Task-specific heads lower retrieval R@1 by 5.93 points (63.35 → 57.42), well above the difference between two training seeds (0.39), and they lower end-to-end R@10m by 1.83 points while adding parameters. One shared head with the same added layer does not show this drop, so the loss comes from separating the two tasks. UnifyGeo is itself a design with task-specific heads and decoders, and GeoSync outperforms it in retrieval after re-ranking (Table R6) and end to end (Table 1). Both results point to the same conclusion: the benefit comes from the jointly learned shared representation.

On reciprocity. GeoSync does show reinforcement between the tasks, and its direction is from fine-grained localization to retrieval: joint learning and LAR improve retrieval (Tables R4 and R7), fine-grained localization accuracy is preserved (2.72 / 1.46 m vs. 2.89 / 1.42 m for the localization-only variant, Table 5), and the end-to-end result improves. We revised this wording in Sec. 4.3 and in the conclusion.

We hope this resolves the reviewer's concerns, and we would be grateful if the reviewer would consider raising the score. We are happy to clarify anything further.

Dear Reviewer W7da,

Thank you again for your review. As the discussion ends soon, we wanted to check whether our response addressed your concerns. We are happy to clarify anything that is still unclear. Thank you!

Dear Reviewer W7da,

Just a quick follow-up before the discussion closes. We hope our response addressed your concerns, and we'd appreciate it if you could consider updating your score. Thank you!

Reviewer_V3EoOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes GeoSync, a unified framework for cross-view image retrieval and fine-grained localization. It learns shared representations for both tasks and performs coarse-to-fine localization using multi-scale feature correlation and residual refinement. Localization evidence is used to re-rank retrieved candidates, while consistency supervision across overlapping satellite tiles improves robustness to off-center query locations. Experiments on VIGOR and DReSS show that GeoSync achieves competitive performance in both retrieval and fine-grained localization and strong end-to-end geolocalization performance.

Critical strengths
  • Most prior work addresses either retrieval or fine-grained localization. An end-to-end solution is indeed needed. The experiments show that localization supervision and localization-aware re-ranking improve retrieval performance, while the shared representation maintains competitive fine-grained localization accuracy.

  • GeoSync achieves strong end-to-end localization performance, outperforming cascaded pipelines that combine dedicated retrieval and fine-grained localization models on both VIGOR and DReSS.

  • The shared representation enables efficient inference by reusing features across retrieval, fine-grained localization, and re-ranking. The reported comparisons show lower latency and peak GPU memory usage than the evaluated cascaded pipelines.

Critical weaknesses
  • The evaluation is limited to panoramic images with known orientation. The method’s robustness to unknown or noisy orientation and its applicability to limited-field-of-view perspective images remain unclear.
Questions for authors
  • Do retrieval and fine-grained localization achieve their best validation performance at the same epoch? Could one task converge earlier and begin to overfit while the other is still improving?

  • Both retrieval and coarse localization compare the same global ground-image descriptor with satellite descriptors. Are the benefits of joint training specific to this shared matching formulation, or do they extend to other fine-grained localization architectures, such as feature correspondence-based methods?

  • Is retrieval or within-tile localization the main bottleneck for the baseline pipelines? Could the authors evaluate all localizers on the same retrieved satellite references to clarify how much of GeoSync’s end-to-end gain comes from better retrieval versus improved robustness to off-center references?

4 comments
AuthorsOpenReview ↗

Response to Reviewer V3Eo

We thank the reviewer for the positive and precise review. The questions led to four new experiments on VIGOR cross-area.

W1. Noisy orientation and limited field of view

Heading noise. We perturbed the heading of the test queries of the paper's model (Table R8).

Table R8: GeoSync under heading error at test time. End-to-end geolocalization.

Heading errorHit Rate ↑R@1m ↑R@5m ↑R@10m ↑
0° (paper)83.6027.2480.3882.56
±10°82.5025.9779.2481.46
±20°79.1823.5275.7178.14
±45°58.6314.8754.8057.60

GeoSync is robust to moderate heading noise: it loses 1.10 R@10m at ±10° and 4.42 at ±20°. Fully unknown orientation is outside the scope of this paper, and we added it to the limitations as future work.

Limited field of view. For a fair comparison, we retrained GeoSync, AuxGeo and HC-Net with the same augmentation (a random 90° to 360° window of the panorama) and tested them on the same images (Table R9).

Table R9: Limited field of view. All methods retrained with the same augmentation. Best in bold.

FoVMethodHit Rate ↑End-to-end R@1m ↑End-to-end R@5m ↑End-to-end R@10m ↑
180°AuxGeo → HC-Net47.698.1536.4140.82
180°GeoSync47.07 → 57.36 with LAR12.7352.4755.74
90°AuxGeo → HC-Net25.992.4816.0619.67
90°GeoSync24.90 → 29.86 with LAR4.6424.6227.75

Before re-ranking, retrieval is on par. LAR still works with a partial view, because it verifies each candidate at its predicted location, and it adds 10.29 Hit Rate points at 180°. Further gains with modules designed for limited field of view are left for future work.

Q1. Do retrieval and fine-grained localization peak at the same epoch?

Yes. Both improve together and plateau over the same epochs, and neither overfits (Table R10). The checkpoint is selected by retrieval R@1, as is standard on VIGOR, and it is within 0.02 m of the best fine-grained localization checkpoint.

Table R10: Both tasks during training. Fine-grained localization with the Primary and Random protocols of Table 7.

Epoch527292 (selected)
Retrieval R@1 ↑61.0262.7163.35
Fine-grained localization Med. (m) ↓, Primary1.501.481.46
Fine-grained localization Med. (m) ↓, Random1.731.691.66

Q2. Is the benefit of joint training specific to the shared matching formulation?

Our results suggest so. We kept the same backbone and losses but gave each task its own descriptor through a separate head, so that retrieval and fine-grained localization no longer match with the same vector. Retrieval R@1 then drops by 5.93 points (63.35 → 57.42), while one shared head with the same added layer shows no such drop (Table R7 in our reply to Reviewer W7da). So the benefit comes from the two tasks sharing the descriptor. We did not train a correspondence-based method jointly with retrieval, so we do not claim that the benefit extends to those architectures. We note it as future work.

Q3. Is retrieval or within-tile localization the bottleneck?

We ran every fine-grained localization method on the same retrieved references, as the reviewer suggested (Table R11).

Table R11: All fine-grained localization methods on the same retrieved references. R@10m. The last two columns split AuxGeo's top-1 references by type. Best in bold.

Fine-grained localizationEnd-to-end, top-1 references of AuxGeoEnd-to-end, top-1 references of GeoSyncBy type of reference: PositiveBy type of reference: Semi-positive
HC-Net70.8078.0696.2677.35
FG²72.2880.0098.4977.81
UnifyGeo71.8279.1297.2080.99
GeoSync74.3782.5698.7594.05
  1. Retrieval is the main bottleneck. With AuxGeo's references, FG² reaches 72.28. With GeoSync's references, the same FG² reaches 80.00. So +7.72 comes from better retrieval alone, and the other baselines gain a similar amount (+7.26 to +7.30), primarily due to our LAR.

  2. Fine-grained localization on semi-positive references is the second one. On the same references, GeoSync is still 2.56 points above FG² (82.56 vs. 80.00). This gap comes only from semi-positive references: on positive ones all methods are above 96%, while on semi-positive ones GeoSync keeps 94.05% and the others drop to 77.35 to 80.99%.

We hope this resolves the reviewer's concerns, and we would be grateful if the reviewer would consider raising the score. We are happy to clarify anything further.

Dear Reviewer V3Eo,

Thank you again for your review. As the discussion ends soon, we wanted to check whether our response addressed your concerns. We are happy to clarify anything that is still unclear. Thank you!

I thank the authors for their clear rebuttal, which answers my question. Most of the findings align with my expectations, though I was somewhat surprised that retrieval and fine-grained localization both peak at the same epoch (there is nothing wrong with it. I guess the peak epoch also very much depend on the loss weight). I will maintain my initial rating.

We thank the reviewer for taking the time to read our response and for the positive feedback. Reviewer is right that the peak epoch also depends on the loss weights (In our setting the fine-grained localization loss is weighted by 0.1 (Appendix A.3)).

Reviewer_z1nhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeoSync, a unified framework for cross-view geolocalization. It integrates satellite image retrieval and fine-grained localization through a shared representation and introduces multi-scale correspondence matching, residual position refinement, spatial consistency supervision, and localization-aware re-ranking.

Critical strengths

GeoSync integrates retrieval and fine-grained localization into a shared representation, which reduces inference overhead.Evaluations on VIGOR and DReSS show improvements in localization accuracy and computational efficiency.

Critical weaknesses
  1. The main components, including multi-scale correlation, residual refinement, contrastive learning, and similarity-based re-ranking, largely build upon established techniques. The contribution is primarily their integration rather than a fundamentally new learning mechanism.
  2. The ablation study removes semi-positive supervision and cross-tile consistency together, making it difficult to determine their individual contributions. Moreover, the improvements may partly depend on the DINOv3 backbone and its fine-tuning strategy.
Questions for authors

1.Could the authors provide separate ablations for semi-positive supervision and cross-tile consistency loss? 2.How much of the performance improvement comes from the proposed framework rather than DINOv3?

3 comments
AuthorsOpenReview ↗

Response to Reviewer z1nh

We thank the reviewer for the feedback. We first clarify the contribution, because it also answers the two questions.

W1. "The contribution is primarily their integration"

GeoSync is not a combination of existing modules. Existing end-to-end systems either cascade a retrieval model with a separate fine-grained localization model, or give each task its own heads and decoder. GeoSync learns both tasks jointly in one shared representation, and this joint learning gives two effects that no combination of existing components provides:

  1. Fine-grained localization reinforces retrieval. LAR verifies each retrieved candidate at its predicted location with the same features, and needs no training. On the retrieval-only variant of Table 5 (same architecture, retrieval loss only), LAR lowers Hit Rate from 76.66 to 62.96. On GeoSync it raises Hit Rate from 75.15 to 83.60. So the gain comes from what the features learn jointly, not from a task-specific module.

  2. Any reference that covers the query becomes usable. When the retrieved reference is semi-positive, GeoSync localizes 94.05% of the queries within 10 m. HC-Net, FG² and UnifyGeo reach 77.35 to 80.99% on the same references.

Q2. How much comes from the framework rather than DINOv3?

Primarily the framework. We put the same DINOv3, trained with the same recipe and fine-tuning strategy, into the strongest cascade. The result barely changes. Training it jointly, as in GeoSync, gives the full gain (Table R1).

Table R1: Backbone vs. framework. End-to-end geolocalization on VIGOR cross-area. Best in bold.

MethodRetrieval backboneMed. (m) ↓R@1m ↑R@5m ↑R@10m ↑
AuxGeo → FG²ConvNeXt-B1.9724.0768.2672.28
DINOv3 retrieval-only → FG²DINOv3 ViT-S+1.9923.8768.1172.56
GeoSyncDINOv3 ViT-S+1.6327.2480.3882.56
  • Backbone alone: +0.28 R@10m. Joint learning on the same backbone: +10.00 R@10m and +12.27 R@5m.
  • We chose DINOv3 ViT-S+ because it is small. GeoSync has 57.4M parameters and takes 6.36 ms per query, against 99.8M and 453.5M parameters and 24.61 ms and 103.10 ms for the two cascades (Table 2).

Q1. Separate ablations for semi-positive supervision and the consistency loss

We report the full ablation, with all four combinations, in Table R2.

Table R2: Separate ablation of semi-positive supervision and the consistency loss. Fine-grained localization with the ground-truth satellite reference, Random reference protocol of Table 7, VIGOR cross-area. Best in bold.

Semi-positive supervisionLcon\mathcal{L}_{\mathrm{con}}Med. (m) ↓R@10m ↑
––8.6852.21
–✓4.8762.86
✓–2.8582.37
✓✓1.6693.54
  • The consistency loss alone reduces the error from 8.68 m to 4.87 m.
  • Semi-positive supervision alone reduces it to 2.85 m.
  • Together they give 1.66 m, so the two components are complementary. This table is now in the paper.

We hope this resolves the reviewer's concerns, and we would be grateful if the reviewer would consider raising the score. We are happy to clarify anything further.

Dear Reviewer z1nh,

Thank you again for your review. As the discussion ends soon, we wanted to check whether our response addressed your concerns. We are happy to clarify anything that is still unclear. Thank you!

Dear Reviewer z1nh,

Just a quick follow-up before the discussion closes. We hope our response addressed your concerns, and we'd appreciate it if you could consider updating your score. Thank you!

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response Summary and Updated Manuscript to All Reviewers

We sincerely thank the reviewers and the area chair for their time and constructive feedback. We ran all the requested experiments. Each reviewer has an individual reply below, and the revised PDF marks all changes in blue (new results in Appendix D). Here we summarize what GeoSync contributes and what the new experiments show.

The contribution of GeoSync is the joint learning of retrieval and fine-grained localization in one shared representation, without separate task-specific representation heads. It is not a new module. This joint learning has three consequences that we now isolate with controlled experiments:

  1. Fine-grained localization reinforces retrieval at almost no cost. LAR verifies every retrieved candidate at its predicted location with the same learned features, without any training. On the retrieval-only variant of Table 5, which has the same architecture, LAR lowers Hit Rate from 76.66 to 62.96. On GeoSync it raises Hit Rate from 75.15 to 83.60, while fine-grained localization accuracy is preserved (Table 5). The gain therefore comes from the jointly learned features.

  2. Fine-grained localization stays accurate on semi-positive references. When the retrieved reference is semi-positive, GeoSync localizes 94.05% of the queries within 10 m, against 77.35 to 80.99% for HC-Net, FG² and UnifyGeo.

  3. The verification score is a reliable confidence signal. It detects wrong retrievals much better than the retrieval similarity (AUROC 0.904 vs. 0.793), which matters for real deployments.

The result is an accurate, small and efficient end-to-end model: 82.56 R@10m against 72.28 for the strongest cascade, with 57.4M parameters and 6.36 ms per query.

New experiments. z1nh: framework vs. backbone, and the separate ablation of semi-positive supervision and the consistency loss. W7da: AuxGeo's PCM used for fine-grained localization, the source of the LAR gain, the same candidate budget as UnifyGeo, and shared readout vs. task-specific heads. V3Eo: heading noise, limited field of view with retrained baselines, training curves, and all fine-grained localization methods on the same retrieved references.

AuthorsOpenReview ↗

Summary of the Discussion

We thank all reviewers for their time and constructive feedback. The reviewers recognized several strengths of our work:

  • A needed end-to-end solution. Reviewer V3Eo noted that "an end-to-end solution is indeed needed", and Reviewer W7da found the focus on off-center references "practically meaningful" and "closer to practical use".
  • A compact and efficient framework. Reviewer z1nh noted that the shared representation "reduces inference overhead", and Reviewer W7da found the framework "compact", with LAR requiring no additional trained re-ranker.
  • Strong end-to-end results. Reviewers z1nh and V3Eo noted the improvements on VIGOR and DReSS, outperforming cascaded pipelines with lower latency and memory.

The main questions in the initial reviews were about where the novelty lies and what drives the gains. We clarified that our contribution is not a new module, but the joint learning of retrieval and fine-grained localization in one shared representation. The new controlled experiments support this:

  • Joint learning, not the backbone (Reviewers z1nh and W7da). With the same DINOv3 backbone and recipe, a cascade barely changes, while joint learning gives the full end-to-end gain. The same LAR hurts retrieval on retrieval-only features and improves it on jointly learned features.
  • Shared representation vs. task-specific heads (Reviewers W7da and V3Eo). Under the same backbone and training, separate task-specific heads lower both retrieval and end-to-end accuracy, while a shared head of the same size does not.
  • Relation to AuxGeo's PCM and UnifyGeo (Reviewer W7da). We clarified the role of MCG and showed that a PCM-style readout alone fails on off-center references. Under the same candidate budget, GeoSync achieves a higher Hit Rate than UnifyGeo, and LAR improves with more candidates at almost no cost.
  • Ablations and robustness (Reviewers z1nh and V3Eo). We reported semi-positive supervision and the consistency loss separately, evaluated all methods on the same retrieved references, and tested heading noise, limited field of view, and training dynamics.

We thank Reviewer V3Eo for engaging in the discussion and for confirming that our response answered their questions. We sincerely thank all reviewers again for their feedback, and we already incorporated all the new results in the revised version of the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.