acceptodds
Under review as a conference paper at ICLR 2027

Beyond Top-1: Geographic Associations for Multi-Reference Cross-View Geo-Localization

Abstract

Large-scale cross-view geo-localization typically follows a retrieval–localization paradigm: candidates are retrieved from an aerial gallery, and the Top-1 reference is selected for local localization. Yet an aerial gallery provides discrete observations of continuous geographic space, and several highly ranked candidates may jointly point to the same local region. Retaining only Top-1 severs geographic associations among candidates and discards complementary spatial localization evidence. We therefore propose GeoHypLoc, a framework for geographically associated multi-reference cross-view localization. It combines query visual support with geographic adjacency in the gallery to organize highly ranked candidates into a positively spatially associated reference set . Preserving each reference's independent georeferencing, it maps their localization evidence into a common real-world geographic space and adaptively aggregates multi-reference evidence throughout coarse-to-fine localization. Finally, it combines region-level visual support with localization-derived spatial evidence to correct the localization response, shifting from single-reference matching to joint constraints over a geographic region. In complete retrieval–localization evaluation on VIGOR, GeoHypLoc improves Acc@3/5/10m over Sample4Geo+CCVPE by 6.60/5.90/4.23 percentage points on Same-area and 2.07/2.14/3.13 percentage points on Cross-area. These results show that organizing multi-reference localization evidence through geographic spatial associations among retrieved candidates can improve large-scale fine-grained cross-view geo-localization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
cross-view geo-localization, geographic spatial association, reference-set organization, multi-reference joint localization, aerial image retrieval.

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_PZvaOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper shows that retrieval-then-localization pipelines in cross-view geolocalization throw away useful information by keeping only the top-1 aerial tile, since neighboring tiles in the top-k often overlap the same area. GeoHypLoc uses off-the-shelf sample4geo to retreive first top-5 images and then it groups the top-5 retrieved tiles by geographic adjacency and keeps the group with the highest retrieval support. It then runs CCVPE (off-the-shelf fine-grained geolocaization model) on each tile in that group, and fuses the outputs on a shared world map at every stage, and refines the final response with two small correction branches. On VIGOR it beats cascaded pipleines (Sample4Geo+CCVPE) on both splits. Gains are clear on same-area and smaller on cross-area.

Critical strengths

1- The motivation is good. The top-5 often contains several overlapping correct image tiles and current pipelines ignore them. So testing the full retrieval-plus-localization pipeline is the right way to evaluate this and give more practical large-scale cross-view geolocalization for real-world applications.

2- The analysis of failure cases is useful. The paper shows that many errors come from ambiguous responses with multiple strong peaks. This gives a clear picture of where current localizers fail and where multiple references help.

Critical weaknesses

1- I’m not convinced the gains come from the proposed idea. In Table 3, just retraining CCVPE on more retrieved tiles and testing on top-1 only (config A) gives almost the same same-area results as the full model. Joint multi-reference localization is even a bit worse than A on same-area, and the geographic grouping doesn’t do better than simple rank-based selection. Since the main claim of the paper is that multi-reference fusion with geographic grouping drives the improvement, this is a big concern. Right now it looks like much of the gain might come from the extra training.

2- The evaluation doesn’t show the method is worth its cost. The only real baseline is Sample4Geo+CCVPE, while stronger localizers like HC-Net, which is more accurate than CCVPE on VIGOR, are missing. UnifyGeo is cited but not compared, even though it’s the closest prior work and also uses multiple retrieved candidates. On top of that, the method roughly doubles latency and uses about seven times more memory than CCVPE for a fairly small gain on cross-area, and everything is tested only on VIGOR with known orientation. Without stronger baselines or a second dataset, it’s hard to tell whether the improvement holds up or justifies the extra compute.

Questions for authors

Clarifications:

1- Was the Sample4Geo+CCVPE baseline fine-tuned with the same 5-epoch schedule and DARK decoding as your model? If not, how much of the gain do you think comes from the extra training and decoding rather than from using multiple references?

2- In Table 3, joint localization (config C) is slightly worse than independent training (config A) on same-area but better on cross-area. How do you explain this, and what does it say about when multi-reference fusion actually helps?

3- How does your method differ from UnifyGeo, which also uses multiple retrieved candidates through reranking? What does your approach offer that UnifyGeo does not?

4- Why did you choose the RNG graph to form H*, rather than simpler options like using all top-5 tiles or grouping tiles within a fixed distance?

5- All results come from a single seed. How stable are they across runs, given that several cross-area differences are only 1 to 2 points?

Suggestions for future work:

1- Comparing against simpler fusion options, such as merging per-tile CCVPE heatmaps on the world map or stitching the tiles into one image, would strengthen the case for the stage-wise fusion design.

2- Testing with stronger localizers such as HC-Net, FG² or Loc², and comparing directly with UnifyGeo, would make the contribution clearer.

3- Results on a second dataset or with unknown orientation would show whether the method generalizes beyond VIGOR with known orientation.

7 comments
AuthorsOpenReview ↗

Overview: revised geographic input, method and manuscript organization

Thank you especially for suggesting final-response fusion and RGB stitching as comparison methods. Your suggestion directly prompted us to reconsider the central contribution: moving from retrieved Top-1 selection and reranking to organizing fine-localization observations through geographic associations among candidates. We conducted the additional experiments, adopted geographically composed RGB input as the main implementation, and revised the method and experimental sections accordingly. The revised method and principal changes are summarized below.

GeoHypLoc uses the geographic relationships among retrieved aerial references to organize the observations supplied to a fine-grained localizer. (i) From the retrieved Top-5 candidates, we use the full-gallery geographic RNG and query-specific visual support to select a spatially coherent reference set H*. (ii) Each reference has its own georeferencing. We resample its RGB observations onto a common geographic canvas, preserving their relative positions and ground sampling scale. At overlapping locations, valid RGB observations are averaged; singly covered locations retain the available observation, and a coverage mask identifies valid support. This geographically composed RGB image and the ground query form one localization sample. (iii) We retain the native coarse-to-fine localization backbone and its basic supervision, add geographic peak and continuous-coordinate objectives for training, and combine regional query visual support with localization evidence to correct the response before continuous coordinate readout.

The geographic peak objective distinguishes support within 1 m of the true position from competing responses beyond 3 m. We implement this formulation with Sample4Geo + CCVPE and with UnifyGeo; the latter retains official retrieval and reranking before reference-set selection.

The revised Section 3.1 describes reference-set selection and geographic input construction; Section 3.2 presents the native localizers and training objectives; Section 3.3 describes response correction and coordinate readout. Section 4 reports the main results, and Section 5 analyzes reference composition, geographic coverage and response correction. Loss ablations are provided in Section 4.3. The original Full interaction implementation, which aligns and exchanges per-reference responses and states during localization, is retained as an alternative implementation in the appendix, together with comparisons against RGB input and final-response Mean fusion.

We also corrected the historical Cross test-ground-truth loading path to use the SliceMatch-corrected labels adopted by CCVPE.

AuthorsOpenReview ↗

Response 1/4: Attribution of gains and the simpler RGB implementation

Your suggestion is valuable for clarifying how geographic observations contribute to localization.

W1, Q1-Q2 and Suggestion 1: training, geographic input and fusion.

The original CCVPE baseline used released weights and its original output, not matched five-epoch fine-tuning and DARK. The independent and joint configurations also differed in readout. We therefore agree that the original table did not isolate the effect of interaction.

We now compare Base, final-response Mean, newly evaluated Full interaction, and geographic RGB input. CC denotes Sample4Geo+CCVPE; Uni denotes UnifyGeo with official retrieval and reranking. Full and RGB use E3 checkpoints; RGB reports mean±sample standard deviation over seeds 17/2026/68. Full reports the individual seed indicated in the table.

SplitMethodSeedAcc@1mAcc@3mAcc@5mAcc@10mMedian (m)
SameCC Basereleased31.68166.97575.90983.7451.618
SameGeoHypLoc (S4G+CC; Mean)1732.11163.51170.91277.6281.671
SameGeoHypLoc (S4G+CC; Full)2336.23071.61579.79386.3111.417
SameGeoHypLoc (S4G+CC; RGB)17/2026/6837.118±0.03171.731±0.05880.058±0.04086.603±0.0461.373±0.003
SameGeoHypLoc (Uni; Mean)1740.42273.84780.64685.1041.259
SameUni Basereleased40.79776.40083.60488.3281.235
SameGeoHypLoc (Uni; Full)1740.86975.21582.01785.9841.242
SameGeoHypLoc (Uni; RGB)17/2026/6842.327±0.11476.932±0.01583.767±0.06387.967±0.0701.193±0.003
CrossCC Basereleased20.84053.47360.78766.3372.557
CrossGeoHypLoc (S4G+CC; Mean)1720.11651.05657.69262.2452.833
CrossGeoHypLoc (S4G+CC; Full)2321.85255.93263.10267.5622.328
CrossGeoHypLoc (S4G+CC; RGB)17/2026/6821.854±0.17255.915±0.22763.294±0.15568.019±0.0892.336±0.022
CrossGeoHypLoc (Uni; Mean)1726.61664.13468.85570.9971.777
CrossUni Basereleased26.42465.33570.50172.6431.768
CrossGeoHypLoc (Uni; Full)1725.91763.62768.23770.1031.806
CrossGeoHypLoc (Uni; RGB)17/2026/6827.484±0.15165.679±0.14870.311±0.10172.358±0.0541.714±0.007

Accuracies are percentages. CC Base uses DARK; Uni Base uses native readout. Mean independently localizes H* members with RGB seed17 E3 weights, aligns and averages their base responses, then applies DARK. Full and RGB include response correction and continuous readout. The RGB rows use the three-seed results of Table 1; Appendix D.3 compares the default single-seed configuration and describes each readout.

Following your suggestion, we adopt geographic RGB input as the main implementation: georeferenced observations are aligned before aerial encoding, valid RGB values are averaged at overlaps, and the native coarse-to-fine localizer processes the resulting region. Full interaction remains an appendix alternative. The evidence supports geographic observation organization without requiring learned stage-wise interaction.

For a matched inference comparison, we also fixed RGB seed17 E3 weights, H* and base-response DARK. RGB Acc@5m is 78.915/60.510% for CC and 83.562/70.313% for Uni (Same/Cross), exceeding the corresponding Mean rows. Mean was not separately trained. See Sec. 3.1 and Appendix D.3.

Cross-label clarification. We use the SliceMatch-corrected labels adopted by CCVPE. Our audit identified a historical Cross evaluation path that still read legacy test GT. Current training and evaluation explicitly bind the corrected labels. With the historical checkpoint fixed, Cross median changed from 2.714 m in the submission to 2.6697 m under the current evaluation.

Q2: interpretation of the original split-dependent result. Independent training already provided substantial Same-area gains. The old A-versus-C comparison changed training composition, test references and readout together; it did not isolate an interaction-induced generalization effect. The revised claim therefore concerns geographic inputs together with training and response processing. Separately, with identical natural K=1 references and DARK, RGB-trained CCVPE improves Acc@5m from 69.946% to 71.082% on Same and 49.537% to 50.184% on Cross: adaptation itself contributes. Fixed-weight member-budget tests and loss/readout analyses distinguish additional effects in Appendices E.5-E.6 and Section 4.3/5.2.

AuthorsOpenReview ↗

Response 2/4: Stronger baselines, UnifyGeo, and measured computational cost

Your suggestion is valuable for strengthening the comparison and separating the sources of improvement.

W2 and Suggestion 2: stronger complete-system baselines.

We added Sample4Geo+HC-Net, +FG² and +Loc², and native UnifyGeo with official retrieval and reranking. The Sample4Geo cascades receive the same retrieved Top-1; UnifyGeo retains its own retrieval path. The following complete retrieval-localization results use corrected labels (Sec. 4.2):

MethodSame: Acc@1 / @5 / @10mCross: Acc@1 / @5 / @10m
Sample4Geo+CCVPE, DARK31.681 / 75.909 / 83.74520.840 / 60.787 / 66.337
GeoHypLoc (S4G+CC)37.118 / 80.058 / 86.60321.854 / 63.294 / 68.019
Sample4Geo+HC-Net38.591 / 81.346 / 86.59321.349 / 65.711 / 69.257
Sample4Geo+FG², no RANSAC39.061 / 82.673 / 89.12823.064 / 66.609 / 70.781
Sample4Geo+Loc², no RANSAC20.966 / 74.587 / 85.73314.075 / 62.289 / 69.277
Native UnifyGeo40.797 / 83.604 / 88.32826.424 / 70.501 / 72.643
GeoHypLoc (Uni)42.327 / 83.767 / 87.96727.484 / 70.311 / 72.358

Accuracies are percentages; GeoHypLoc rows are three-seed means. Loc² excludes 2/4 failed Same/Cross queries (N=52,603/53,690); other rows use 52,605/53,694 queries. Baselines use released weights, not matched E3 fine-tuning. The direct-cascade improvement motivates applying geographic organization within the stronger UnifyGeo pipeline.

Q3: what is added beyond UnifyGeo?

UnifyGeo already uses retrieval and fine matching together and reranks candidates. We retain this capability: official global retrieval -> official reranking -> H* selection -> geographically organized RGB localization. Our additional step uses independently georeferenced candidate footprints to form one localization observation in a common geographic coordinate system, rather than passing the reranked first reference alone to localization. Its native four-stage localization path is adapted to this geographic input, with the training constraints and response processing in Secs. 3.2–3.3.

Relative to native UnifyGeo, Same/Cross Acc@1m increases by 1.530/1.060 percentage points and median error falls from 1.235/1.768 m to 1.193/1.714 m. Acc@10m decreases by 0.361/0.286 points. These gains support the design goal of improving fine localization precision. The small loss-ablation differences are reported separately in Section 4.3.

W2: is the improvement worth the additional computation?

We replaced concurrent-run estimates with isolated measurements on one RTX 4090, using 258 fixed queries per split, warmup, repeated measurements and matched batching within each backbone. Entries below are single-reference -> RGB:

Backbone / splitSerial localization (ms/query)Peak PyTorch allocated (GiB)
CCVPE / Same47.915 -> 87.3400.385 -> 1.108
CCVPE / Cross46.963 -> 77.2000.385 -> 1.030
UnifyGeo / Same30.295 -> 49.1200.311 -> 0.516
UnifyGeo / Cross29.582 -> 44.2230.311 -> 0.480

Serial localization includes warm-cache image processing, device transfer, model computation and coordinate readback. It starts from cached candidates and excludes full-gallery retrieval and reranking. RGB/single-reference serial-time ratios are 1.823/1.644 for CCVPE and 1.621/1.495 for UnifyGeo. CUDA-synchronized forward-only ratios are higher: 3.246/3.127 and 2.279/2.196, respectively. Allocated memory is not total GPU-process memory. Sampled process peaks are 1.164 GiB for CCVPE Top-1 versus 4.297/3.512 GiB for RGB (Same/Cross), and 0.795 GiB for UnifyGeo Top-1 versus 1.316/1.213 GiB for RGB. Appendix F also reports reserved-memory peaks.

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

This paper introduces GeoHypLoc, a framework for multi-reference cross-view geo-localization that challenges the conventional practice of performing fine-grained localization using only the Top-1 retrieved aerial image. The key motivation is that geographically adjacent aerial references may contain complementary information about the same ground-level location, which is discarded when retrieval results are reduced to a single reference.

Critical strengths
  1. A meaningful problem formulation that exploits geographic structure in retrieved references. The strongest conceptual contribution is recognizing that aerial retrieval candidates are not independent samples but spatially related observations of a continuous geographic region.
  2. Encouraging end-to-end improvements supported by detailed diagnostic analyses.
Critical weaknesses
  1. The experiments do not convincingly establish that the proposed geographic joint-localization mechanism is responsible for the main improvements.

  2. Limited generalization evidence and an unfavorable computational trade-off. The experimental validation is restricted to VIGOR, a fixed Sample4Geo retriever, Top-5 candidates, and known orientation. There is no evidence that the same benefits persist with different retrieval models, different gallery densities, or different candidate budgets.

Questions for authors
  1. How does GeoHypLoc compare with simple, geographically aligned multi-reference inference baselines under matched training and decoding protocols?

  2. How much of the improvement over the original CCVPE baseline is attributable to retraining, changed training data composition, response correction, and coordinate decoding?

  3. How sensitive are the results to reference selection and geographic graph construction?

6 comments
AuthorsOpenReview ↗

Overview: revised geographic input, method and manuscript organization

Thank you for recognizing the geographic formulation and for your suggestions on attribution, controlled comparisons and generalization. To address these questions, and inspired by another reviewer’s suggestion to evaluate RGB stitching, we refocused the paper on organizing fine-localization observations through geographic associations among retrieved candidates. The additional experiments led us to adopt the simpler geographic RGB input as the main implementation, retain the original Full interaction implementation as an appendix comparison, and analyze the sources of improvement through training, loss and readout experiments. We first outline the revised method, then address your questions individually.

GeoHypLoc uses the geographic relationships among retrieved aerial references to organize the observations supplied to a fine-grained localizer. (i) From the retrieved Top-5 candidates, we use the full-gallery geographic RNG and query-specific visual support to select a spatially coherent reference set H*. (ii) Each reference has its own georeferencing. We resample its RGB observations onto a common geographic canvas, preserving their relative positions and ground sampling scale. At overlapping locations, valid RGB observations are averaged; singly covered locations retain the available observation, and a coverage mask identifies valid support. This geographically composed RGB image and the ground query form one localization sample. (iii) We retain the native coarse-to-fine localization backbone and its basic supervision, add geographic peak and continuous-coordinate objectives for training, and combine regional query visual support with localization evidence to correct the response before continuous coordinate readout.

The geographic peak objective distinguishes support within 1 m of the true position from competing responses beyond 3 m. We implement this formulation with Sample4Geo + CCVPE and with UnifyGeo; the latter retains official retrieval and reranking before reference-set selection.

The revised Section 3.1 describes reference-set selection and geographic input construction; Section 3.2 presents the native localizers and training objectives; Section 3.3 describes response correction and coordinate readout. Section 4 reports the main results, and Section 5 analyzes reference composition, geographic coverage and response correction. Loss ablations are provided in Section 4.3. The original Full interaction implementation, which aligns and exchanges per-reference responses and states during localization, is retained as an alternative implementation in the appendix, together with comparisons against RGB input and final-response Mean fusion.

We also corrected the historical Cross test-ground-truth loading path to use the SliceMatch-corrected labels adopted by CCVPE.

AuthorsOpenReview ↗

Response 1/4: Revised implementation and geographically aligned baselines

Your suggestion is valuable for clarifying how geographic observations contribute to localization.

W1 and Q1: matched baselines and the revised geographic input.

The original CCVPE baseline used released weights and its original output, not matched five-epoch fine-tuning and DARK. The independent and joint configurations also differed in readout. We therefore agree that the original table did not isolate the effect of interaction.

We now compare Base, final-response Mean, newly evaluated Full interaction, and geographic RGB input. CC denotes Sample4Geo+CCVPE; Uni denotes UnifyGeo with official retrieval and reranking. Full and RGB use E3 checkpoints; RGB reports mean±sample standard deviation over seeds 17/2026/68. Full reports the individual seed indicated in the table.

SplitMethodSeedAcc@1mAcc@3mAcc@5mAcc@10mMedian (m)
SameCC Basereleased31.68166.97575.90983.7451.618
SameGeoHypLoc (S4G+CC; Mean)1732.11163.51170.91277.6281.671
SameGeoHypLoc (S4G+CC; Full)2336.23071.61579.79386.3111.417
SameGeoHypLoc (S4G+CC; RGB)17/2026/6837.118±0.03171.731±0.05880.058±0.04086.603±0.0461.373±0.003
SameGeoHypLoc (Uni; Mean)1740.42273.84780.64685.1041.259
SameUni Basereleased40.79776.40083.60488.3281.235
SameGeoHypLoc (Uni; Full)1740.86975.21582.01785.9841.242
SameGeoHypLoc (Uni; RGB)17/2026/6842.327±0.11476.932±0.01583.767±0.06387.967±0.0701.193±0.003
CrossCC Basereleased20.84053.47360.78766.3372.557
CrossGeoHypLoc (S4G+CC; Mean)1720.11651.05657.69262.2452.833
CrossGeoHypLoc (S4G+CC; Full)2321.85255.93263.10267.5622.328
CrossGeoHypLoc (S4G+CC; RGB)17/2026/6821.854±0.17255.915±0.22763.294±0.15568.019±0.0892.336±0.022
CrossGeoHypLoc (Uni; Mean)1726.61664.13468.85570.9971.777
CrossUni Basereleased26.42465.33570.50172.6431.768
CrossGeoHypLoc (Uni; Full)1725.91763.62768.23770.1031.806
CrossGeoHypLoc (Uni; RGB)17/2026/6827.484±0.15165.679±0.14870.311±0.10172.358±0.0541.714±0.007

Accuracies are percentages. CC Base uses DARK; Uni Base uses native readout. Mean independently localizes H* members with RGB seed17 E3 weights, aligns and averages their base responses, then applies DARK. Full and RGB include response correction and continuous readout. The RGB rows use the three-seed results of Table 1; Appendix D.3 compares the default single-seed configuration and describes each readout.

Based on the additional experiments, we adopt geographic RGB input as the main implementation: georeferenced observations are aligned before aerial encoding, valid RGB values are averaged at overlaps, and the native coarse-to-fine localizer processes the resulting region. Full interaction remains an appendix alternative. The evidence supports geographic observation organization without requiring learned stage-wise interaction.

For a matched inference comparison, we also fixed RGB seed17 E3 weights, H* and base-response DARK. RGB Acc@5m is 78.915/60.510% for CC and 83.562/70.313% for Uni (Same/Cross), exceeding the corresponding Mean rows. Mean was not separately trained. See Sec. 3.1 and Appendix D.3.

Cross-label clarification. We use the SliceMatch-corrected labels adopted by CCVPE. Our audit identified a historical Cross evaluation path that still read legacy test GT. Current training and evaluation explicitly bind the corrected labels. With the historical checkpoint fixed, Cross median changed from 2.714 m in the submission to 2.6697 m under the current evaluation.

Interpreting the original comparisons. Independent training already provided substantial Same-area gains. The old A-versus-C comparison changed training composition, test references and readout together; it did not isolate an interaction-induced generalization effect. The revised claim therefore concerns geographic inputs together with training and response processing. Separately, with identical natural K=1 references and DARK, RGB-trained CCVPE improves Acc@5m from 69.946% to 71.082% on Same and 49.537% to 50.184% on Cross: adaptation itself contributes. Fixed-weight member-budget tests and loss/readout analyses distinguish additional effects in Appendices E.5-E.6 and Section 4.3/5.2.

AuthorsOpenReview ↗

Response 2/4: Separating training, response correction and coordinate readout

Your suggestion is valuable for strengthening the comparison and separating the sources of improvement.

Q2: how much comes from retraining, training composition, response correction and decoding?

Training and reference composition. The released CCVPE baseline was not fine-tuned with our schedule. In the original independent-training comparison, rank-matched references and H* gave similar Same-area results, so those results do not establish a distinct advantage from H*-based training alone. The revised RGB system combines geographic input organization with training adaptation. It is not a test-only modification of official weights.

To examine adaptation without extra test references, we use naturally single-member H* queries, preserving the same query, retrieved image and GT for both models. Both use base-response DARK without correction:

Backbone / splitNOfficial: Acc@1 / @5mRGB-trained: Acc@1 / @5m
CCVPE / Same11,44631.749 / 69.94634.353 / 71.082
CCVPE / Cross23,33218.177 / 49.53718.490 / 50.184
UnifyGeo / Same15,29439.329 / 76.57339.395 / 76.435
UnifyGeo / Cross27,25123.698 / 59.48423.970 / 59.246

Accuracies are percentages. CCVPE benefits even without multiple test references, while UnifyGeo changes are smaller and mixed. Appendix E.6 gives the paired protocol. The fixed-RGB-weight Mean comparison in Response 1 instead isolates how the references are used during inference.

Added training objectives. We perform a 2x2 peak/coordinate-loss ablation, retaining the same inputs, architecture and final readout within each backbone. Enabling both rather than neither changes Acc@5/10m by +0.981/+1.163 points on CCVPE Same and +2.375/+3.112 on Cross. Its Acc@1m changes by -0.065/-0.106 points. UnifyGeo's four accuracy changes are only +0.044/+0.087/+0.082/+0.124 points on Same and +0.024/+0.058/+0.019/+0.091 on Cross. We therefore treat these as supporting objectives, not the sole source of the method's improvement. We interpret the smaller UnifyGeo gains as related to its existing joint retrieval-localization representation. See Section 4.3.

Response correction at fixed decoding. We fix seed17 E3 weights, H*, queries and DARK, changing only the decoded response from Z0 to corrected Z. Strong-secondary queries are defined before correction by a secondary/primary peak ratio of at least 0.25, with the secondary peak at least 2 m away:

Backbone / splitStrong-secondary share (%)Correction gain: Acc@1 / @5m (points)
CCVPE / Same20.827+0.767 / +2.355
CCVPE / Cross34.324+1.655 / +5.670
UnifyGeo / Same9.389+0.223 / +0.952
UnifyGeo / Cross13.350+0.112 / +0.558

The gains concentrate in queries with competing responses. See Sec. 5.2 and Appendix E.1.

Coordinate readout. DARK is a decoding operation, not a training loss. Continuous readout is evaluated separately after correction. For CCVPE Strong-secondary, moving from corrected DARK to continuous readout changes Same/Cross Acc@1m from 11.336/8.367% to 10.250/6.799%, while Acc@10m rises from 59.620/41.595% to 62.422/43.142%. Thus response correction and coordinate readout have distinct precision/success trade-offs. Appendix E.1 reports the individual readouts.

Reviewer_4uYhOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes to find geographically coherent regions in a top-k ground-aerial retrieval list, noting that the aerial tiles are not independent - they are crops of a continuous surface. Instead of taking the top-1 results for fine localization, the authors show that finding neighboring regions in the subset and localizing there improves accuracy. First, they show that such continous regions are abundant in retrieval on VIGOR: more than 50% of queries contain at least 2 positive, neighboring aerial tiles in its top-5 predictions. They propose to find the largest set via relative neighborhood and visual similarity weights. Following CCVPE, independent localization responses are iteratively aligned and fused based on position, and a joint localization likelihood distribution is estimated over the merged area. An evidence-guided response correction updates the localization heatmap based on visual support. The final location is estimated with DARK. Evaluations on VIGOR show that this substantially improves localization accuracy over CCVPE, especially in the same-area split.

Critical strengths
  • The idea of finding geographically adjacent sets in the top-k retrieval list is novel and interesting. I agree with the authors that top-1 retrieval is suboptimal, and they suggest a technically sound alternative by fusing geographically consistent aerial tiles in a retrieval shortlist, and show that this indeed improves results. They support this design decision with clear numbers, such as the frequency how often multiple, neighboring aerial tiles appear in typical top-k retrieval lists.
  • The approach of fusing information and responses from neighboring cells is technically sound. The authors build on CCVPE, but make distinct design choices to fuse information from different aerial tiles based on global position. Experiments show that this indeed helps localization accuracy.
  • The appendix is very detailed and informative.
  • Visual examples are of high quality and support the proposed method.
Critical weaknesses
  • In VIGOR aerial tiles are overlapping, while in other scenarios they are not. It remains unclear to what extent overlapping tiles are a necessity in the proposed work. Other works use non-overlapping aerial tiles (Scaling GeoLoc, Lindenberger et al., NeurIPS2025), and it would be interesting how well the proposed method would work in such scenarios.
  • Evaluation is performed on a single dataset (VIGOR), questioning generalization. From Table 1, one can already see a sharp decline of the proposed methods effectiveness between same and cross split. An evaluation on even more out-of-distribution data (like OpenCVL, Xia et al., 2026, or KITTI) is missing.
  • The differentiation to CCVPE is unclear from the main paper and without reading CCVPE. While the method section highlights structural differences especially on the fusion of different aerial tiles, a clear A-B comparison highlighting the differences would help to isolate the contributions.
  • Related work is in the appendix, and I am not sure whether this is allowed by the template. As this section highlights the differentiation to literature, I would recommend moving this to the main paper.
  • Many important baselines are missing. The paper mainly compares to Sample4Geo and CCVPE, but SotA baselines on VIGOR, especially under the known orientation setting, are missing (e.g. Loc2, Xia et al., ICLR 2026). While these methods only solve top-1 localization, they could easily be paired with Sample4Geo top-1 retrieval, top-5 reranking based on the number of inlier matches. This would strengthen the work, and also answer whether the proposed work outperforms top-k re-ranking directly.
Questions for authors
  • What is the difference between overlap state map and overlap response map? In Figure 2 they look identical.
  • How would the proposed method perform under top-10 or top-100 retrieval stitching? Would this improve results?
  • Some notations are confusing, e.g. the different number of decimals in Tab. 2.
  • What is the intuition that the Acc@1cm drops in the cross-area split?
  • If no geographically consistent subset >= 2 tiles exist, does the method maintain CCVPE accuracy?
7 comments
AuthorsOpenReview ↗

Overview: revised geographic input, method and manuscript organization

Thank you for recognizing the geographic formulation, diagnostic evidence and visual examples. Your suggestions on overlap conditions, the distinction from CCVPE, stronger baselines and a second dataset helped us improve the methodological explanation and evaluation. In parallel, another reviewer’s suggestion to evaluate RGB stitching prompted us to simplify the implementation while retaining the central theme of geographically organizing fine-localization observations: references are composed into an RGB input according to their geographic positions and processed by the native localizer. The original Full interaction implementation is retained as an appendix comparison. We outline the method and manuscript changes below before addressing your questions.

GeoHypLoc uses the geographic relationships among retrieved aerial references to organize the observations supplied to a fine-grained localizer. (i) From the retrieved Top-5 candidates, we use the full-gallery geographic RNG and query-specific visual support to select a spatially coherent reference set H*. (ii) Each reference has its own georeferencing. We resample its RGB observations onto a common geographic canvas, preserving their relative positions and ground sampling scale. At overlapping locations, valid RGB observations are averaged; singly covered locations retain the available observation, and a coverage mask identifies valid support. This geographically composed RGB image and the ground query form one localization sample. (iii) We retain the native coarse-to-fine localization backbone and its basic supervision, add geographic peak and continuous-coordinate objectives for training, and combine regional query visual support with localization evidence to correct the response before continuous coordinate readout.

The geographic peak objective distinguishes support within 1 m of the true position from competing responses beyond 3 m. We implement this formulation with Sample4Geo + CCVPE and with UnifyGeo; the latter retains official retrieval and reranking before reference-set selection.

The revised Section 3.1 describes reference-set selection and geographic input construction; Section 3.2 presents the native localizers and training objectives; Section 3.3 describes response correction and coordinate readout. Section 4 reports the main results, and Section 5 analyzes reference composition, geographic coverage and response correction. Loss ablations are provided in Section 4.3. The original Full interaction implementation, which aligns and exchanges per-reference responses and states during localization, is retained as an alternative implementation in the appendix, together with comparisons against RGB input and final-response Mean fusion.

We also corrected the historical Cross test-ground-truth loading path to use the SliceMatch-corrected labels adopted by CCVPE.

AuthorsOpenReview ↗

Response 1: overlap, single-reference cases and larger candidate sets

W1: dependence on overlapping references.

Your suggestion is a valuable addition to testing the geographic conditions under which the method works. We quantified VIGOR overlap before evaluating reduced reference availability.

Within the fixed local diagnostic region, the area covered by multiple references accounts for 99.27%/99.92% on Same/Cross. Removing all three labeled semi-positive references reduces it to 81.76%/86.67%, and the mean number of references covering GT falls from 3.814/4.040 to 1.073/1.141. No local zero-overlap query was observed under this intervention. VIGOR therefore remains overlapping even after removing the labeled semi-positive references.

Within the rebuttal period, we could not reconstruct a fully non-overlapping benchmark. We instead evaluated this reduced-reference condition with fixed models, repeating retrieval and H* selection. CCVPE RGB Acc@10m changes from 86.655% to 83.543% on Same and from 67.918% to 64.102% on Cross; UnifyGeo RGB changes from 87.895% to 86.049% and from 72.420% to 69.630%. We consider overlapping reference support important to the observed multi-reference benefit: reducing that support lowers localization accuracy, while substantial accuracy remains. The overlap statistics and localization diagnostic are described in Appendix E.4.

Q5: what happens when no multi-member subset exists?

We evaluated naturally occurring |H*|=1 queries with identical query images, retrieved references and GT. Both the official model and RGB-trained model use base-response DARK decoding without response correction:

Backbone / splitNOfficial Acc@1 / @5mRGB-trained Acc@1 / @5m
CCVPE / Same11,44631.749 / 69.94634.353 / 71.082
CCVPE / Cross23,33218.177 / 49.53718.490 / 50.184
UnifyGeo / Same15,29439.329 / 76.57339.395 / 76.435
UnifyGeo / Cross27,25123.698 / 59.48423.970 / 59.246

CCVPE retains its single-reference accuracy: Acc@1/5m improves by 2.604/1.136 points on Same and 0.313/0.647 points on Cross. Thus, the geographic-input training does not sacrifice these single-member results when a multi-reference region is unavailable. UnifyGeo Acc@1m also improves slightly. See Appendix E.6.

Q2: Top-10 and Top-100.

Your suggestion usefully tests whether a larger retrieval budget supplies additional geographic evidence. We completed RNG/MASS statistics for these candidate budgets:

SplitTop-kMean / maximum H* membersH* AnyPositive (%)
Same52.327 / 594.521
Same102.887 / 1094.898
Same1005.616 / 5695.533
Cross51.847 / 575.860
Cross102.264 / 1076.308
Cross1004.632 / 5777.267

From Top-5 to Top-100, selected-H* AnyPositive increases by only 1.011/1.408 percentage points, while mean membership more than doubles. PositiveFraction, the mean proportion of officially positive (Strong or semi-positive) references within each selected H*, decreases from 87.044% to 63.040% on Same and from 71.718% to 55.679% on Cross. Thus, larger budgets mainly expand the selected input with proportionally fewer positive references. We consider the limited additional selected-region coverage, together with the larger input, a practical justification for Top-5.

AuthorsOpenReview ↗

Response 2: CCVPE comparison, state versus response, and presentation

W3: comparison with CCVPE.

Your suggestion is valuable for making the contribution clear without requiring readers to study CCVPE first. In the revised RGB implementation, we retain CCVPE's native coarse-to-fine localization backbone and do not add a member-interaction architecture. The change is to its observation input: a geographically registered H* RGB region replaces the single reference. Native supervision is retained, with geographic peak and coordinate objectives, followed by regional-support response correction. Secs. 3.1–3.3 describe these steps. The structure/function comparison is therefore now much simpler.

Q1: overlap state map versus overlap response map.

In the original Full implementation, the response map contains scalar matching evidence at geographic positions. The state map contains multi-channel intermediate localization features passed to subsequent stages. They share geographic coordinates and a coverage mask but have different contents and roles. Position-dependent member weights aggregate both; the fused state then affects later matching. At singly covered positions, the valid member receives weight one. Their similar appearance in the original diagram did not communicate this distinction clearly.

The revised RGB main implementation composes observations before aerial encoding and has no separate cross-member state-fusion module. Appendix G gives the original Full state/response formulation and recurrence; Appendix D.3 compares its localization results with RGB and final-response Mean.

W4: related work in the main paper.

Thank you for this helpful suggestion on clarifying the paper's positioning. We have placed Related Work in Section 2 of the main paper. It discusses cross-view retrieval, fine-grained localization and unified retrieval–localization, including UnifyGeo and the added localization baselines. The introduction also explains the distinction from HLoc, Set-CVGL and Slice-Loc: GeoHypLoc organizes independently georeferenced aerial observations after retrieval. Training-loss ablations are presented in Section 4.3; detailed implementations remain in the appendix.

Q3: inconsistent precision in Table 2.

We standardized the displayed precision and units. The correct-reference tables report Mean and Median in metres, separately for known and unknown orientation. Published results and our reevaluations are identified, and the Loc² failed-query exclusions are stated. See Sec. 4.2 and Appendix D.5.

Q4: Cross tight-threshold results.

Your observation helped us identify an important evaluation inconsistency. The historical Cross test path read the earlier GT coordinates; we corrected it to the SliceMatch labels adopted by CCVPE. At fixed checkpoint and queries, the verified Cross median changes from 2.714 m to 2.6697 m.

The revised RGB main results use this corrected evaluation. Cross Acc@1m is 21.854±0.172% versus 20.840% for the CCVPE cascade, and 27.484±0.151% versus 26.424% for native UnifyGeo. These are the current model comparisons, separate from the fixed-checkpoint label correction. As discussed in Secs. 4.2 and 5, geographic observation organization improves fine localization, while response correction uses regional query support to resolve competing spatial evidence.

Comments on the paper

1 comment
Reviewer_gdASOpenReview ↗

Thank you for the author's reply. I will maintain my original opinion.

Thank you for the author's reply. I will maintain my original opinion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.