- 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
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 K | UnifyGeo | GeoSync |
|---|---|---|
| No re-ranking | 61.17 / 68.24 | 63.35 / 75.15 |
| 5 (UnifyGeo's default) | 67.58 / 72.27 | 67.90 / 82.18 |
| 10 | 67.47 / 72.24 | 68.72 / 83.16 |
| 20 | 67.20 / 71.92 | 69.09 / 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.
| Readout | Extra params | Retrieval: R@1 / Hit | Retrieval: after LAR | End-to-end Med. ↓ | End-to-end R@1m ↑ | End-to-end R@5m ↑ | End-to-end R@10m ↑ |
|---|---|---|---|---|---|---|---|
| Shared descriptor (GeoSync) | none | 63.35 / 75.15 | 69.09 / 83.60 | 1.63 | 27.24 | 80.38 | 82.56 |
| One shared head (control) | +0.30M | 62.60 / 74.38 | 68.85 / 83.56 | 1.64 | 27.15 | 80.26 | 82.54 |
| Two task-specific heads | +0.59M | 57.42 / 69.58 | 65.72 / 81.45 | 1.80 | 25.92 | 78.41 | 80.73 |
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!
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:
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.
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.
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.
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.