acceptodds
Under review as a conference paper at ICLR 2027

GeoCoRe: Recovering Missing Cross-View Relations under Disjoint Pairwise Supervision for Tri-View Geo-Localization

Abstract

Tri-view geo-localization integrates UAV, satellite, and ground views, with fully supervised settings typically relying on direct paired supervision for the cross-view relations of interest. Such supervision is difficult to obtain when geospatial data are independently collected across platforms or datasets, leaving some relations unobserved during training. This work considers an incomplete tri-view setting by recovering the UAV–ground (u–g) relation from UAV–satellite (u–s) and satellite–ground (s–g) pairs, without direct u–g supervision or shared location instances between the two observed relations. Existing pairwise methods can learn both observed relations effectively yet still struggle to recover the missing u–g relation. Because the two observed relations are supervised on disjoint location sets, each pairwise objective constrains only its own relation, and the shared satellite view alone does not establish cross-edge comparability. This motivates organizing the two relations under shared coordinates and a common metric. GeoCoRe is introduced as a two-stage framework for recovering missing cross-view relations in tri-view geo-localization. Stage 1 induces shared Geo-primitives as a shared primitive reference, while Stage 2 uses this reference to guide coordinate assignment and align shared-coordinate representations across the observed edges. GeoCoRe improves the average missing-edge R@1 from 1.53% for the strongest baseline to 6.11%, demonstrating the value of cross-relation compositionality for recovering unobserved relations. Code is available at https://anonymous.4open.science/r/GeoCoRe.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Cross-view geo-localization, multi-view representation learning, incomplete pairwise supervision, composable representation learning, missing-relation recovery

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_V25WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies UAV–ground retrieval without direct UAV–ground training pairs. It trains on UAV–satellite and satellite–ground pairs from disjoint locations. GeoCoRe uses frozen DINOv3 features, shared learnable queries, and local spatial-relation modeling in Stage 1. Stage 2 uses the query responses to guide optimal-transport feature aggregation. Both stages train on the two available relations; at test time, UAV and ground descriptors are compared directly. The main experiment trains on DenseUAV and VIGOR and reports 6.11% average UAV–ground Recall@1 on University-1652.

Critical strengths
  1. The setting poses a useful generalization question. Independently collected UAV–satellite and satellite–ground datasets can cover different locations. Testing whether they support retrieval on a relation absent from training supervision is a reasonable way to study transfer between view pairs.

  2. The reported improvements are not confined to DINOv3. Table 1 includes five backbones. With ConvNeXt-B, GeoCoRe improves average UAV–ground Recall@1 over Sample4Geo from 1.53% to 2.00%; with DINOv2 ViT-B/14, it improves over LKGL from 0.60% to 5.58%. These results provide evidence of an advantage beyond simply selecting DINOv3.

  3. The ablations and same-dataset experiment provide useful evidence. Table 2 improves from 3.94% with Stage 1 alone to 5.42% with unguided shared aggregation and descriptor combination, and to 6.11% with primitive guidance. Table 3 supports sharing projection heads. In the disjoint-pair experiment within University-1652, the full model achieves 10.83%, compared with 7.20% for Pairwise and 7.44% for Stage 1. The gains therefore persist without the main experiment's cross-dataset shift.

Critical weaknesses
  1. Excluding all direct UAV–ground training pairs needs stronger motivation. The introduction cites independent data collection and scarce complete tri-view correspondences. These make incomplete supervision plausible. However, direct UAV–ground training requires only two views, and paired training data already exist in University-1652. The paper needs to explain what concrete data limitation makes such pairs unavailable or inadequate, and why training only on the two indirect relations is useful under those conditions. Excluding direct supervision defines a valid generalization experiment, but does not by itself establish the practical need for this training restriction.

  2. The presentation obscures substantial overlap with prior work and overstates the novelty of the aggregation mechanisms. In Stage 1, the shared primitive reference begins with eight learnable queries that aggregate patch features through cross-attention. This basic mechanism is established in retrieval methods such as BoQ (Ali-bey et al., CVPR 2024). Sharing query parameters does not ensure that the same query attends to corresponding content across views. GeoCoRe adds local spatial-relation modeling and consistency constraints; the contribution needs to be identified at that level.

    Stage 2's learned assignments, Sinkhorn transport, dustbin, and grouping-based feature aggregation closely overlap with SALAD: Optimal Transport Aggregation for Visual Place Recognition (Izquierdo and Civera, CVPR 2024), which is neither cited nor discussed. Calling these operations “shared-coordinate mapping” obscures this relationship. The primitive-based reference, occupancy priors, and assignment constraints appear to be incremental modifications to this established aggregation pattern. The ablations demonstrate a benefit within the proposed architecture, but do not establish a substantial advance over a competitive shared aggregator such as SALAD. Together with the unacknowledged methodological overlap, this leaves Stage 2's claimed contribution insufficiently substantiated for acceptance.

  3. The experiments show better retrieval on the missing relation, but do not establish the proposed explanation for that improvement. Both stages learn from the two observed relations, and the final descriptors are directly compared on UAV–ground images. A shared SALAD model trained on the same relations would also permit this evaluation. The key question is what the proposed guidance adds beyond strong shared aggregation. Table 2 supports an increment of 0.69 percentage points over the authors' unguided variant, but this is not a comparison with SALAD: the variant uses a different mass allocation.

    Appendix A correctly shows that pairwise similarity constraints alone do not determine cross-relation similarities, while acknowledging that shared encoders and pretraining restrict the ambiguity. It does not show that this ambiguity is the main cause of errors in the actual model. Likewise, sharing query or aggregation indices does not establish content correspondence across views. Tables 2–3 and Figure 4 support the design, but do not establish that it improves retrieval specifically by correcting “coordinate incompatibility.” A matched shared aggregation baseline is needed to assess the proposed mechanism's additional value.

Questions for authors
  1. Training-data restriction: Under what concrete data-acquisition conditions are direct UAV–ground training pairs unavailable or inadequate? Why do existing paired datasets not meet that need, making the proposed indirect training setup useful?

  2. Actual methodological contribution: Please properly cite and discuss SALAD and relate the shared-query design to prior attention-based aggregation. Which specific changes beyond these established mechanisms are essential for the reported gains? What evidence shows that shared primitives capture corresponding content across views, beyond using the same query indices?

  3. Shared aggregation comparison: How does a shared SALAD model perform when trained on the same two relations with a matched backbone, descriptor size, and training budget? To isolate Stage 2's contribution, replacing its guided aggregation with SALAD while retaining the Stage 1 features and descriptor combination would also be informative. These comparisons would clarify the value of primitive guidance beyond ordinary shared representation learning.

5 comments
AuthorsOpenReview ↗

Official Response to Reviewer V25W [1/3]

W1 & Q1. Practical Motivation for Missing UAV–Ground Supervision

1. Overview

Independently collected u–s and s–g data may cover different locations, making it impossible to construct direct same-location u–g supervision. Existing u–g paired data also cannot automatically fill the supervision gap at these locations. GeoCoRe aims to recover the missing relation from existing two-view data, reducing dependence on additional paired data collection and annotation.

We thank the reviewer for highlighting these practical data constraints. We further clarify the specific source of this supervision gap and the coverage limitations of existing datasets.

(1) Pairing-coverage limitations arising from independent data collection and existing datasets.

UAV imagery may form u–s pairs with satellite images, whereas street-view and satellite images collected at other locations may form s–g pairs. When the locations in the two datasets do not overlap, obtaining direct u–g supervision requires additional collection of the missing view and verification of the correspondences.

Existing datasets reflect this difference in pairing coverage. DenseUAV and SUES-200 primarily provide u–s pairs, while VIGOR, CVUSA, and CVACT primarily provide s–g pairs. Although University-1652 contains direct u–g pairs, it covers different locations and thus cannot directly supply the missing same-location correspondences in the above data sources. Introducing these direct pairs as additional cross-domain supervision would represent a different data-availability condition.

(2) Indirect supervision provides useful learning signals for the missing relation.

As shown in Table 6 of manuscript, in the same-domain, location-disjoint experiment on University-1652, directly using pretrained DINOv3 features yields an RmissR_{\mathrm{miss}} of 2.67%, Pairwise, trained using only u–s and s–g supervision, reaches 7.20%, and GeoCoRe further improves it to 10.83%. This indicates that indirectly observed relations can provide useful learning signals for missing u–g retrieval, and that GeoCoRe can use these signals more effectively to improve missing-relation recovery.

2. Revision

In the revised Introduction, we will clarify the pairing-coverage limitations caused by independent data collection and use examples from existing datasets to explain the practical motivation for incomplete relation supervision.

AuthorsOpenReview ↗

Official Response to Reviewer V25W [2/3]

W2 & W3 & Q2 & Q3. Methodological Contribution and Cross-Edge Comparability

1. Overview

GeoCoRe's core contribution is to use a shared primitive reference for primitive-guided coordinate assignment and to improve the cross-edge comparability of independently supervised relations under a common metric. New experiments show that GeoCoRe improves RmissR_{\mathrm{miss}} by 2.00 pp and 1.45 pp over Shared SALAD and Stage 1 + SALAD, respectively, demonstrating additional gains beyond shared aggregation.

We thank the reviewer for pointing out the technical connections between GeoCoRe and BoQ/SALAD. In the revision, we will add citations to these methods and clarify the differences between established aggregation mechanisms and GeoCoRe. To further assess this methodological contribution, we add two SALAD comparisons mentioned by the reviewer and combine them with the existing ablations, primitive visualizations, and a new representation analysis to provide further evidence.

2. Experiments and Analysis

(1) Methodological Distinction: From general shared aggregation to cross-edge comparable representations.

BoQ uses learnable queries and cross-attention to aggregate local features. SALAD constructs global retrieval descriptors through SALAD assignment implemented with Sinkhorn-based optimal transport (OT), a dustbin, and local-feature aggregation. These methods can learn discriminative representations, but shared queries or aggregation parameters alone do not guarantee sufficient cross-edge comparability between two relations supervised by independent datasets.

GeoCoRe addresses this issue by a coordinated design. Stage 1 combines local spatial-relation modeling with cross-view primitive consistency constraints to learn a shared primitive reference. Stage 2 uses primitive responses to construct a primitive-guided reference assignment and an image-level occupancy prior. Finally, shared local projection and coordinate fusion functions establish a common metric.

Therefore, the key distinction is that shared primitive responses are explicitly converted into the basis for primitive-guided coordinate assignment across relations, while a common metric improves the comparability of independently supervised relations. This design serves missing-relation recovery rather than merely enhancing feature aggregation for a single observed relation.

(2) SALAD Comparisons: Assessing the additional value of the complete design over general shared aggregation.

Following the reviewer's suggestion, we add two comparisons, Shared SALAD and Stage 1 + SALAD. The former trains a shared SALAD aggregator on the two observed edges. The latter retains the same frozen Stage 1, base descriptor, and fusion scheme, while replacing only the guided aggregation in Stage 2 with SALAD. All methods use the same disjoint-location u–s/s–g supervision, a frozen DINOv3 ViT-B/16 backbone, the same input resolution, and a 1536-dimensional final descriptor. They are evaluated on University-1652 using the same u–g Query and full Gallery.

Table R10. Comparison with shared SALAD aggregation under the cross-dataset missing-edge protocol.

Methodu2g R@1g2u R@1RmissR_{\mathrm{miss}}u2g mAPg2u mAP
Shared SALAD3.954.274.113.552.87
Stage 1 + SALAD4.464.854.663.783.18
Full GeoCoRe5.986.246.114.934.01

As shown in Table R10, when the Stage 1 representations and descriptor-fusion scheme are held fixed, the complete Stage 2 still achieves better recovery performance, supporting the additional value of Stage 2 over standard SALAD aggregation.

AuthorsOpenReview ↗

Official Response to Reviewer V25W [3/3]

(3) Cross-Edge Comparability: Assessing the roles of the shared reference, coordinate organization, and common metric.

Shared primitive indices do not automatically guarantee cross-view content correspondence. The key to GeoCoRe is that primitive responses guide the subsequent coordinate assignment, while learning over the two observed edges improves the comparability of cross-edge representations.

Tables 2–3 in the manuscript demonstrate the complementary roles of the validated shared Primitive Reference, Primitive-Guided Coordinate Assignment and Dual-Edge Metric Alignment. Notably, upon removal of the unified metric, the performance for observed edges improved whilst that for missing edges deteriorated, indicating that the discriminatory power of observed relationships does not necessarily translate into comparability across relationships.

Figure 5 in Appendix E shows local responses at the same primitive indices across the three views and illustrates how these responses contribute to constructing the primitive-guided reference assignment QvQ^v and the final content-adaptive coordinate assignment PvP^v.

To further analyse the impact of Primitive-guided Coordinate Mapping on cross-view representation, we compare the location discrimination capabilities of Stage 1 Geo-primitives and Stage 2 Shared Coordinates on the University-1652 dataset. Specifically, the average cosine similarity of cross-view features is calculated separately for the same location and for different locations, and the difference between the two is used to measure the ability to distinguish between locations.

Table R11. Cross-view place discriminability of same-index representations. Values denote the cosine-similarity margin between matched and unmatched locations for each view pair.

Representationu–ss–gu–g (Missing)
Stage 1 Geo-primitive representations0.002850.000680.00057
Stage 2 Shared-Coordinate Representations0.3320.1570.153

Table R11 shows that, following Stage 2 coordinate mapping, the co-indexed representations of the three viewpoint relationships all exhibit stronger ability to distinguish locations, including the u–g relationship, which was not directly supervised. Taken in conjunction with the comparison with Stage 1 + SALAD in Table R10, these results support the contribution of Primitive-guided Coordinate Mapping to the comparability of cross-viewpoint representations in terms of both retrieval performance and representational distinctiveness.

3. Revision

In the revised manuscript, we will add citations to BoQ and SALAD and clarify their technical differences from GeoCoRe. We will also add the controlled SALAD comparisons and cross-view representation analysis to Appendices D and E.

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

This paper proposes a weakly supervised two-stage framework for the Tri-view Geo-localization task. In the absence of UAV–ground image alignment supervision, it uses satellite imagery as an intermediary to learn a set of shared local feature representations for tri-view retrieval. The proposed framework achieves competitive experimental performance on the constructed benchmark.

Critical strengths

The task setting is practically meaningful. Joint tri-view (UAV–Satellite–Ground) retrieval is closer to the real needs of low-altitude navigation and emergency response than conventional two-view matching, and the weakly supervised setting further reduces the annotation burden.

The method design is effective. Using satellite imagery as a cross-view intermediary to bridge UAV and ground views is a clear and well-motivated strategy, and the experimental results show considerable performance gains on the retrieval task.

Critical weaknesses

W1. The ecological validity of the benchmark is questionable; low absolute performance may partly reflect synthetic test views and imprecise cross-view correspondence rather than task difficulty alone.

The constructed benchmark is based in part on University-1652 for evaluation. It should be noted that University-1652 is used here only as an evaluation source, not as training supervision; the concern is therefore not about non-realistic training signals, but about the validity of the benchmark as an evaluation set. The UAV images in University-1652 are not field-captured but are rendered/synthesized using the Google Earth 3D engine. Evaluating on such synthetic UAV views may reward representations that exploit the projection regularities and scene-geometry priors of the rendering engine, rather than genuine viewpoint changes caused by altitude, pose, lighting, and occlusion in real UAV flights. At the same time, the ground views in University-1652 are also essentially synthetic and sparse; subsequent work has explicitly pointed out that the dataset contains only about 3.38 ground images per location on average, and all of them are synthetic samples. Moreover, the UAV–ground correspondences are not sufficiently precise. Therefore, the low absolute performance reported on this benchmark should not be attributed simply to the difficulty of the task itself; the synthetic nature of the test UAV views and the sparsity/imprecision of the UAV–ground correspondences are also very likely important factors constraining the measured performance. Without discussing this test-side bottleneck, the claim that the benchmark is “realistic” lacks sufficient support.

W2. The gain at the backbone level needs to be re-confirmed after controlling variables.

The net gain of the method relative to the backbone has not yet been sufficiently isolated. Specifically, DMNIL also uses ConvNeXt-T as its backbone and achieves higher performance than the proposed method on observed edges. DMNIL’s own ablation experiments further show that an untrained ConvNeXt-Tiny achieves only 9.55% R@1 on UAV–satellite retrieval, whereas introducing its learning mechanism raises this to 71.06%, indicating that the discriminative power of the backbone itself is limited and that the performance differences mainly come from the training strategy rather than the feature extractor. In this situation, comparing final metrics only across different methods cannot rule out confounding effects from backbone choice or pretraining differences. It is necessary to re-evaluate the actual gain of the proposed method over competing methods under controlled conditions with a unified backbone and unified input resolution.

Questions for authors

Real-world validity of the benchmark. The benchmark relies on U1652, whose UAV views are 3D-rendered rather than field-captured and whose UAV–ground correspondences are sparse/synthetic. Does the submission include any evaluation on real UAV imagery, e.g., the real UAV subset of GeoLoc/GeoBridge or the real portion of UniGeoRS? If not, how do the authors justify the claim that the benchmark is realistic, and how do they distinguish low absolute performance caused by task difficulty from that caused by synthetic/sparse cross-view supervision?

Matched-backbone comparison with baselines. The paper reports the proposed method across several backbones, but the competing methods are not evaluated under the same backbones. Can the authors provide matched-backbone results for DMNIL and the strongest baselines under ConvNeXt-T and DINO at the same input resolution, reporting “backbone only,” “backbone + proposed training strategy,” and “full method”? In particular, since DMNIL with ConvNeXt-T outperforms the proposed method on observed edges, can the authors clarify whether the reported advantage over baselines stems from the proposed training strategy or from backbone choice?

2 comments
AuthorsOpenReview ↗

Official Response to Reviewer otkA [1/2]

W1 & Q1. Benchmark Validity and Generalization to Real Aerial Imagery

1. Overview

The ground images in University-1652 are collected from real street-view imagery rather than generated synthetically. New zero-shot experiments show that GeoCoRe continues to substantially outperform the strongest comparison method, Sample4Geo, on an independent real-aerial dataset, indicating that its advantage in missing-relation recovery is not limited to the synthetic UAV view.

We thank the reviewer for raising the issue of evaluation realism. We clarify that the UAV images in University-1652 were generated using Google Earth 3D, whereas the ground images are real street-view images. The original paper (Zheng et al., 2020) reports an average of 3.38 street-view images per location.

With regard to GeoLoc, GeoBridge and UniGeoRS, as mentioned by the reviewers, it is not currently possible to obtain three-view data and annotations that would support a full evaluation under the same protocol. Therefore, we retain the reproducible University-1652 evaluation and add a zero-shot retrieval experiment on real aerial imagery from CVOGL to assess the method on a non-synthetic aerial view.

2. Experiments and Analysis

(1) We construct an independent Aerial–Ground retrieval evaluation using real aerial imagery.

Using the original cross-view object annotations from CVOGL (Sun et al., 2023), we construct 742 positive real aerial–ground retrieval pairs without using model predictions or visual similarity to establish correspondences. GeoCoRe and Sample4Geo use identical Query and Gallery sets and the same positive-pair definitions. We directly evaluate their existing weights, with no additional training, fine-tuning, or model selection.

Table R6. Zero-shot Aerial–Ground retrieval on real aerial imagery from CVOGL.

MethodA2G R@1G2A R@1RmissR_{\mathrm{miss}}A2G mAPG2A mAP
Sample4Geo-Joint4.997.556.2711.3013.33
Full GeoCoRe10.389.039.7018.1316.93

(2) Testing on real aerial has further validated the recovery capabilities across different test domains.

GeoCoRe increases RmissR_{\mathrm{miss}} from 6.27% to 9.70%, with higher mAP in both directions. This result indicates that GeoCoRe's advantage in missing-relation retrieval is not specific to the synthetic UAV view in University-1652, but extends to zero-shot matching between real aerial and ground images in an independent evaluation.

3. Revision

In the revised manuscript, we will clarify the image sources and location-level evaluation protocol of University-1652. We will also add the construction of the CVOGL cross-view correspondences, the Query/Gallery settings, and the zero-shot retrieval results to Appendix D.

AuthorsOpenReview ↗

Official Response to Reviewer otkA [2/2]

W2 & Q2. Matched-Backbone Comparison and Analysis of Method Gains

1. Overview

With matched backbones, GeoCoRe consistently outperforms the baselines on missing-edge retrieval using both ConvNeXt-T and DINOv3 ViT-B/16. Additional experiments show that GeoCoRe's core design provides additional gains beyond those from the pretrained backbone and training strategy.

We thank the reviewer for raising the issues of backbone fairness and the sources of the performance gains. We add a matched-backbone method comparison and layered controls covering Backbone-only, the training strategy, and Full GeoCoRe to examine the sources of the gains.

2. Experiments and Analysis

(1) Matched-backbone comparison: GeoCoRe retains its missing-edge recovery advantage across different visual backbones.

We re-evaluate GeoCoRe and the competing methods with ConvNeXt-T and DINOv3 ViT-B/16. Within each backbone group, all methods use the same pretrained initialization and input resolution. All task-trained methods use disjoint-location DenseUAV u–s and VIGOR s–g supervision, without direct u–g pairs, and are evaluated using the same Query set and full Gallery on University-1652.

Table R7. Matched-backbone comparison under the cross-dataset missing-edge protocol.

BackboneMethodU2GG2URobsR_{\mathrm{obs}}RmissR_{\mathrm{miss}}
ConvNeXt-TBackbone-only0.380.7010.190.54
DMNIL-Joint0.600.8523.620.73
Full GeoCoRe1.841.4318.241.64
DINOv3 ViT-B/16DMNIL-Joint2.813.2220.883.02
Sample4Geo-Joint3.584.6524.404.12
No-Core (Backbone + Proposed Training Strategy)3.705.0419.934.37
Full GeoCoRe5.986.2430.926.11

With matched backbones, GeoCoRe achieves better missing-edge retrieval performance with both backbones. In particular, with ConvNeXt-T, DMNIL achieves higher observed-edge performance but lower missing-edge performance, further indicating that observed-edge discriminability does not necessarily translate into missing-edge recovery capability.

(2) Layered gain analysis: disentangling the contributions of pretrained features, the training strategy, and the complete method design.

We conduct two complementary evaluations. On University-1652, we compare Backbone-only, Backbone + Proposed Training Strategy, and Full GeoCoRe under location-disjoint supervision to distinguish the gains from task training and the full method (Table R8). Since University-1652 uses narrower-FoV Ground views than the full panoramas used for training on VIGOR, we further evaluate these variants on VIGOR panoramas to examine the standalone retrieval capability of DINOv3 and whether GeoCoRe provides additional gains without relying on the change in Ground FoV (Table R9).

Table R8. Disentangling backbone, training-strategy, and method gains on University-1652.

MethodRobsR_{\mathrm{obs}}U2GG2URmissR_{\mathrm{miss}}
Backbone-only20.862.063.282.67
Backbone + Proposed Training Strategy34.567.227.177.20
Full GeoCoRe42.5910.3011.3610.83

These experiments further show that our core design contributes additional gains beyond the base training strategy. This result is consistent with the No-Core experiment under the original cross-dataset protocol (Table R7), further supporting GeoCoRe's additional methodological gains under different supervision conditions.

Table R9. DINOv3 ViT-B/16 retrieval performance on VIGOR using full panoramic ground images.

MethodG2S R@1S2G R@1
DINOv3 ViT-B/16 Backbone-only0.600.81
No-Core (Backbone + Proposed Training Strategy)2.494.95
Full GeoCoRe4.267.87

The VIGOR full-panorama experiment provides further evidence that GeoCoRe offers additional gains over the base training strategy.

3. Revision

In Appendix D of the revised manuscript, we will add the complete six-direction retrieval results for the matched-backbone comparison and layered gain experiments, and specify the pretrained initialization, input resolution, data splits, and training configurations. We will also add retrieval analyses using different forms of ground-view input.

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

This work studied an under-explored and practically motivated missing-edge CVGL task, where model needs to estimate the correspondence between UAV and ground images with trained only on disjoint u-s and s-g pairs. It proposed GeoCoRe, which exploits shared learnable primitive queries and coordinate assignment to represent different views in a common latent coordinate system. The experiments report improvements over adapted baselines, together with component ablations and a within-University-1652 control.

Critical strengths
  • This work provides a meaningful task formulation. The missing-edge setting captures a plausible gap in available paired data and asks a question broader than ordinary two-view geo-localization.
  • It provides a coherent method and demonstrates improvements over existing methods. These results support the utility of the method under the tested protocols, while the ablations begin to identify useful design choices.
Critical weaknesses
  • The feasibility limits of the missing-edge task is not well established. All methods in table one (including the proposed one) have low absolute missing-edge R@1 (e.g., CeoCoRe gets 5.98 for U2G). Without a reference evaluated on the same test queries and gallery, it is difficult to judge how much of the remaining error comes from the missing training edge, from limited shared visual evidence in the test images, or from transfer across datasets and scene types. In fact, the two observed edges alone do not ensure that u and g retain the same place-specific cues. A model trained with direct u-g supervision on separate University-1652 samples and evaluated on the same u–g test set in the paper would provide a useful feasibility reference.
  • The scope of the edge-completion remains narrow. The evaluation focus on recovering u-g from disjoint u-s and s-g. We need to also investigate cases of recovering s-u and s-g to get a full comprehension of the proposed task.
  • The adapted-baseline comparison needs an original task performance check. Several prior works are only designed for two-view case. It adapting them to the three-view setting may potentially cause model collapse. If adaptation substantially degrades that ability, the missing-edge comparison becomes less informative. So we need to check the performance of those methods in their original settings after adapting to the three-view task.
  • The contribution of the complex processing pipeline is not fully isolated. From the paper, my personal hypothesis is that the use of a set of shared primitive queries could have played a central role in view-alignment, as this methods is widely used under the multi-modal setting. Can we isolate the contribution of those queries? also, it remains unclear how much the later coordinate reassignment adds beyond a simple readout of Stage 1 primitives.
Questions for authors
  • Could you report the a direct-U–G-supervised reference, trained on separate University identities and evaluated with the same test galleries and a comparable backbone and training budget?
  • What evidence distinguishes insufficient u/g place information from a learned alignment that fails to transfer across scene or feature types?
  • Can the same protocol be run with a different missing relation, e.g., s–u inferred from s–g and g–u within University-1652?
  • Could you check, report, and discuss the performance of the adapted baselines on their original protocol?
  • How does the full method compare with (a) a model retrained without primitive queries, maybe using a capacity-matched readout of patch features TvT^v, and (b) a simpler readout that replaces the Stage 2 pooled slot features zqvz_q^v with a dimension-matched cross-attention outputs Z~v\tilde Z^v from Stage 1 (we need to train the fusion head using Z~v\tilde Z^v with keeping Stage 1 frozen)?
6 comments
AuthorsOpenReview ↗

Official Response to Reviewer cvMZ [1/4]

W1 & Q1 & Q2. Feasibility and Limitations of Missing u–g Recovery

1. Overview

The u–g relation contains learnable information. Independent Pairwise learning does not fully exploit this information, but GeoCoRe can effectively improve missing-relation recovery without direct u–g supervision.

First, we clarify that this paper does not assume that the two observed relations inherently guarantee recovery of the missing relation. Rather, this is the central question investigated in this work: under disjoint-location u–s and s–g supervision, how can cross-edge comparability be established to improve retrieval on the unobserved u–g relation?

To address the reviewer's questions regarding feasibility and sources of error, we add Direct U–G Supervised Reference, and analyze it together with the existing in-domain experiment, cross-edge ablations, and cross-dataset results.

2. Experiments and Analysis

(1) Direct supervision verifies learnability, while the in-domain comparison evaluates missing-relation recovery.

Following the reviewer's suggestion, we split the 701 official training locations of University-1652 into 560 training locations and 141 validation locations and train Direct U–G Reference. This experiment uses the same backbone, input resolution, and training budget as GeoCoRe.

Table R1. Directly supervised feasibility reference and missing-edge u–g retrieval on University-1652. All methods use identical test queries and galleries, with disjoint training and test identities.

MethodTraining SupervisionU2G R@1G2U R@1RmissR_{\mathrm{miss}}
Direct U–G ReferenceDirect u–g10.5511.2810.91
PairwiseDisjoint u–s / s–g7.227.177.20
GeoCoReDisjoint u–s / s–g10.3011.3610.83

Direct U–G Reference verifies that the task is learnable, Pairwise reveals the limitations of learning the two relations independently, and GeoCoRe improves RmissR_{\mathrm{miss}} from 7.20% to 10.83% under the same missing-supervision setting (+3.63 pp). Furthermore, GeoCoRe achieved results close to those of the Direct U–G Reference (10.91%) without direct U–G supervision. The latter serves as an empirical feasibility reference rather than a performance upper bound.

(2) The cross-edge ablation verifies the role of a common metric in missing-edge recovery.

Table 3 of the manuscript shows that removing Dual-Edge Metric Alignment increases RobsR_{\mathrm{obs}} from 30.92% to 32.83%, while decreasing RmissR_{\mathrm{miss}} from 6.11% to 3.94%.

This opposing trend shows that improved observed-edge retrieval performance does not necessarily translate into better missing-edge recovery. Even when effective discriminative features are learned independently for the two observed relations, the absence of a common metric weakens cross-edge comparability and consequently limits missing-edge recovery.

(3) Cross-dataset experiments evaluate the applicability of missing-relation recovery.

In the original DenseUAV u–s + VIGOR s–g → University-1652 setting, the training and test data also differ in scene distribution and ground-image format. GeoCoRe still substantially outperforms Sample4Geo (6.11% vs. 1.53%), indicating that the proposed mechanism for establishing cross-edge comparability remains effective under compound domain shifts.

3. Revision

In the revised manuscript, we will add the experimental setting and results of Direct U–G Reference and use the in-domain University-1652 experiment together with Table 3 to further clarify the learnability of the missing relation and the role of cross-edge comparability.

AuthorsOpenReview ↗

Official Response to Reviewer cvMZ [2/4]

W2 & Q3. Generality of Missing-Edge Recovery

1. Overview

GeoCoRe's recovery advantage is not limited to the u–g relation or to satellite as the intermediate view. The new experiments improve performance in both missing-u–s and missing-s–g settings, demonstrating applicability to different missing relations.

We thank the reviewer for suggesting further evaluation on other missing relations. We add two groups of experiments that, together with the original setting, cover all three single-missing-edge configurations in the three-view relation graph.

2. Experiments and Analysis

(1) Evaluation of alternative missing edges under disjoint-location supervision.

We build two additional configurations on University-1652:

  • Setting A: s–g + g–u → u–s.
  • Setting B: u–s + u–g → s–g.

The 701 locations in the official training set are divided into two disjoint subsets of 350 and 351 locations for the two observed relations, and 35 locations in each group were retained for validation. The missing relationships do not participate in training, validation, or model selection. All methods use the same data split, training strategy, and model selection criterion and are evaluated on the same official test Query and full Gallery.

Table R2. Recovery of alternative missing relations on University-1652. Both settings follow disjoint-location supervision without missing-edge pairs. Results are R@1 (%).

SettingMethodMissing Direction 1Missing Direction 2RobsR_{\mathrm{obs}}RmissR_{\mathrm{miss}}
A: Missing u–sSample4GeoU2S: 47.95S2U: 57.635.5052.79
GeoCoRe50.2467.6210.2658.93
B: Missing s–gSample4GeoG2S: 1.40S2G: 2.2837.911.84
GeoCoRe4.236.7046.355.47

(2) Both new missing relationships show consistent recovery advantages.

GeoCoRe improves RmissR_{\mathrm{miss}} by 6.14 pp and 3.63 pp under the missing u–s and missing s–g settings, respectively, and outperforms Sample4Geo in all four missing relation retrieval directions.

Combined with the original u–g experiment (+4.58 pp), GeoCoRe shows a recovery advantage in all three single missing-edge configurations, indicating that its effectiveness does not depend on a specific missing relation or the Satellite intermediate perspective.

3. Revision

In Appendix D of the revised manuscript, we will add the complete protocols and six-direction retrieval results for these new experiments.

AuthorsOpenReview ↗

Official Response to Reviewer cvMZ [3/4]

W3 & Q4. Adapted-Baseline Validity and Comparison Fairness

1. Overview

The adapted baselines retain retrieval capability on their original two-view tasks, while GeoCoRe achieves better observed-edge retrieval under the unified protocol. The additional internal comparison with No-Core provides independent methodological evidence for GeoCoRe's missing-edge recovery gain.

We thank the reviewer for raising concerns about the validity of the baseline adaptations and the fairness of the comparisons. It is important to clarify that the original methods were primarily optimized for their respective two-view datasets and retrieval objectives, but this paper studies recovery of the missing u–g relation under joint, disjoint-location u–s/s–g supervision. Because the training objectives and supervision conditions differ, the fairness of the adapted comparison cannot be judged directly from the specialized training performance reported in the original papers.

Therefore, we add an evaluation on the original two-view tasks and an internal comparison with No-Core to assess, respectively, the retrieval capability of the adapted models and the additional methodological gain of GeoCoRe.

2. Experiments and Analysis

(1) Original-task evaluation: the adapted models retain two-view retrieval capability.

First, we clarify that our baseline adaptation modifies the training-supervision protocol rather than redesigning the core network architecture of each method. We retain each method's original feature extraction and aggregation mechanisms and only unify the supervision datasets as joint training on disjoint-location DenseUAV u–s and VIGOR s–g data.

Specifically, the original Sample4Geo is trained for g–s retrieval on VIGOR, and the original DMNIL is trained for u–s retrieval on DenseUAV. After introducing another heterogeneous observed relation, the models must jointly optimize two retrieval objectives, and the resulting joint-training performance may differ from that obtained via the original single-edge training. Consequently, performance differences across relations for the adapted models reflect their learning and adaptation under the unified supervision setting.

To directly address the reviewer's request regarding original-task performance, we evaluate the jointly trained adapted checkpoints under each method's native Query/Gallery protocol.

Table R3. Original-task retrieval performance of adapted baselines under their native Query/Gallery protocols. † denotes evaluation on a dataset not used by the original method.

MethodDenseUAV U2S R@1DenseUAV S2U R@1VIGOR G2S R@1VIGOR S2G R@1
Sample4Geo-Joint60.66†56.93†3.076.65
DMNIL-Joint9.2210.550.044†0.045†
GeoCoRe67.2875.324.267.87

The adapted baselines retain retrieval capability on their original tasks. Sample4Geo-Joint achieves 3.07%/6.65% R@1 for VIGOR g→s/s→g, while DMNIL-Joint achieves 9.22%/10.55% R@1 for DenseUAV u→s/s→u. GeoCoRe achieved higher R@1 across all four observation directions, demonstrating the advantages of its joint learning of the u–s and s–g relationships.

(2) Internal comparison with No-Core: Verifying the independence of performance gains from missing edges.

To further exclude differences in external baseline adaptation from the assessment of the method's gain, we construct an internal No-Core comparison using the same DINOv3 backbone, data split, and two-observed-edge training strategy. No-Core retains the DINOv3 backbone and the corresponding joint-training strategy over the two observed edges but does not include the GeoCoRe modules.

Table R4. Missing-edge retrieval under the unified DINOv3 setting.

MethodRmissR_{\mathrm{miss}}
Sample4Geo-Joint (DINOv3)4.12
DMNIL-Joint (DINOv3)3.02
No-Core (DINOv3 + Joint Training)4.37
Full GeoCoRe (DINOv3)6.11

GeoCoRe improves RmissR_{\mathrm{miss}} from 4.37% for No-Core to 6.11% (+1.74 pp) and also outperforms the backbone-matched Sample4Geo-Joint and DMNIL-Joint configurations. This internal comparison shows that GeoCoRe's core design contributes an independent missing-edge recovery gain under the unified training setting.

3. Revision

In Appendix D of the revised manuscript, we will add the unified observed-edge retrieval results and the internal comparison, together with detailed descriptions of the training settings, data splits, and Query/Gallery evaluation protocols.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

I. Acknowledgments

We thank all reviewers for their careful evaluation of our manuscript and their constructive suggestions. In response to their concerns regarding the research motivation, methodological novelty, experimental fairness, and effectiveness of missing-relation recovery, we conducted several additional ablation studies and validations across data conditions. Below, we summarize the main findings concerning the paper's central research question. Detailed experimental protocols and results are provided in our responses to the respective reviewers.

II. Problem, Key Insight and Methodological Contribution

Research Problem. When UAV–Satellite and Satellite–Ground data are independently collected by different platforms and cover disjoint locations, the available supervision cannot directly provide UAV–Ground training pairs for corresponding locations. We investigate how to recover the UAV–Ground relation, which is entirely unobserved during training, using only these two observed relations.

Key Insight. Effectively learning the two observed relations does not guarantee that they can be composed into a reliable missing relation. Existing Pairwise methods can learn matching along the observed edges, but their representations across relations may not be sufficiently comparable. The shared-metric ablation in the main text further shows that missing-edge performance can instead decrease as observed-edge performance improves, validating this key concern.

Methodological Contribution. GeoCoRe addresses cross-relation representation comparability through an interlinked two-stage design: Stage 1 learns a shared primitive reference, and Stage 2 uses primitive responses for primitive-guided coordinate assignment and optimizes representations from different observed relations under a shared metric. The core contribution is to make the shared primitive reference the basis for organizing representations across relations, thereby improving missing-relation recovery, rather than to introduce a new cross-attention or OT aggregation operator.

III. Key Concerns and New Evidence

Key QuestionsReviewersNew Evidence and Conclusions
1. Is the missing-edge problem practically motivated and learnable?v25W, cvMZIndependently collected datasets reflect the limited location coverage of dual-view supervision. The added Direct u–g Reference experiment verifies that u–g correspondence is learnable. In the in-domain experiment, GeoCoRe achieves 10.83% without direct u–g supervision, outperforming Pairwise (7.20%) and approaching the directly supervised u–g reference (10.91%).
2. What is new beyond established shared aggregation?v25W, cvMZWe clarify the technical differences from BoQ and SALAD. Relative to Shared SALAD and Stage 1 + SALAD, which retains the same Stage 1, GeoCoRe improves performance by 2.00 pp and 1.45 pp, respectively. The Query-free and Stage 1 + Simple Readout ablations, as well as cross-view representation analysis, further support the contributions of the shared primitive reference and subsequent coordinate mapping.
3. Are the gains independent of backbone choice and training strategy?otkA, cvMZWe add matched comparisons using ConvNeXt-T and DINOv3 and evaluate the adapted baseline under the original Query/Gallery protocol. Under the same DINOv3 backbone and dual-observation edge training conditions, GeoCoRe improves RmissR_{\mathrm{miss}} from 4.37% to 6.11% compared to Backbone + Proposed Training Strategy, validating the additional gains of the complete method.
4. Does recovery generalize beyond the original missing relation?cvMZWe add the Missing u–s and Missing s–g settings, in which RmissR_{\mathrm{miss}} improves by 6.14 and 3.63 pp, respectively. When combined with the original u–g experiment, these results show that GeoCoRe provides a recovery advantage across all three single-missing-edge configurations.
5. Does the advantage extend beyond synthetic UAV evaluation views?otkAWe add zero-shot Aerial–Ground retrieval on real aerial imagery from CVOGL, increasing RmissR_{\mathrm{miss}} from 6.27% to 9.70%. This result indicates that the advantage in missing-relation recovery extends to an independent test set of non-rendered aerial imagery.

IV. Commitment to Revision

We will clarify the motivation for missing supervision and distinguish GeoCoRe from BoQ and SALAD, while adding matched-backbone, stage-ablation, missing-edge, and real-aerial results in Appendices D–E.

Overall, this new evidence supports the paper's central conclusion: effectively learning independent observed relations does not automatically ensure composability across relations, whereas primitive-guided coordinate assignment and a shared metric can further improve missing-relation recovery and maintain this advantage across different missing relations and test-data conditions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.