acceptodds
Under review as a conference paper at ICLR 2027

WideUAV: Cross-View Satellite-Tile Retrieval and Target-Extent Grounding

Abstract

In wide-area satellite imagery, finding the matching tile is only the first step toward localizing a target observed from a UAV: the target must also be grounded within that tile. We study cross-view geo-grounding, a unified setting that combines satellite-tile retrieval with target grounding. To support this task, We introduce WideUAV, a rendered benchmark comprising 3,505 wide satellite tiles annotated with off-centered target bounding boxes and 112,160 UAV queries captured under controlled variations in altitude and heading. In contrast to existing cross-view benchmarks such as University-1652, where satellite images are tightly cropped to 120m × 120m around the target, WideUAV uses 2370m × 1328m satellite images, covering over 200× the area. The drone-view queries are sampled at 32 views per location on a fixed grid of 4 altitudes and 8 headings. We further propose UNISAG (Unified Satellite Retrieval and Grounding), a single model that shares visual representations across retrieval and grounding while incorporating UAV heading and altitude cues. To jointly evaluate retrieval and grounding, we introduce retrieval-gated uIoU, an end-to-end metric that assigns zero uIoU when the correct satellite tile is not retrieved. On this metric, UNISAG achieves 22.67%, outper- forming several competitive re-implemented approaches, including TransGeo, with an 18.87-point gain over the strongest baseline at 3.80%. We also observe an asymmetric interaction between the two parts: grounding supervision improves the retrieval representation, whereas retrieval supervision improve little to grounding.

Primary area:
datasets and benchmarks
Keywords:
Retrieval, Cross-view localization, UAV benchmark

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7ipCOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces WideUAV, a benchmark for cross-view satellite-tile retrieval and target-extent grounding, comprising 3,505 satellite tiles and 112,160 UAV images. It combines wide satellite coverage with off-center targets and controlled variations in UAV altitude and heading. The authors also propose UNISAG, which integrates retrieval and grounding through shared visual representations and incorporates altitude and heading information for localization.

Critical strengths

WideUAV addresses a meaningful problem by combining satellite-tile retrieval with within-tile target localization over wide satellite regions with off-center targets. The proposed benchmark and accompanying method provide a useful foundation for future research.

Critical weaknesses

The rationale for the altitude-dependent box sizes and the practical assumptions behind gallery construction need further clarification. In addition, using the same rendering source for UAV and satellite images leaves uncertainty about how well the results generalize to real UAV imagery.

Questions for authors
  1. Appendix A.4 maps UAV altitudes of 150, 200, 250, and 300 m to nominal box side lengths of 330, 414, 528, and 678 pixels, respectively. Could the authors clarify how these specific box sizes were determined?

  2. The main evaluation provides each query with 100 candidate tiles, including its paired tile. In practical use, how would the geographic coverage of the satellite gallery be determined before localizing the query? Would this require prior knowledge of the operating area or an approximate UAV location? Does the method assume that the target is always within the predefined gallery?

  3. Both the UAV and satellite images are generated using the same Google Earth rendering pipeline. Could the authors discuss whether this shared source reduces the difficulty of cross-view matching, and what implications it has for applying the method to real UAV imagery?

4 comments

We thank the reviewer for the constructive feedback and thoughtful suggestions. The revised manuscript has been uploaded, with all revisions highlighted in blue for ease of reference. Below, we provide point-by-point responses to each comment.

Q1: The rationale for the altitude-dependent box sizes needs further clarification. Appendix A.4 maps UAV altitudes of 150, 200, 250, and 300 m to nominal box side lengths of 330, 414, 528, and 678 pixels, respectively. Could the authors clarify how these specific box sizes were determined?

A1: The bounding-box centers are determined from the target coordinates in the satellite images, while their side lengths are obtained through manual calibration of the common target region observed across UAV views, rather than directly calculated from altitude. Specifically, for each UAV altitude, we first select four UAV views with headings of 0∘0^\circ, 90∘90^\circ, 180∘180^\circ, and 270∘270^\circ. In each view, we identify a ground reference point near the bottom-center pixel (1920,2160)(1920,2160) of the 3840×21603840\times2160 UAV image and locate its corresponding position in the north-up satellite image. These four reference positions define the boundaries of the target region in the satellite image. We then manually calibrate a square bounding box to enclose this region, obtaining nominal side lengths of 330, 414, 528, and 678 pixels for altitudes of 150, 200, 250, and 300 m, respectively.

Since the camera intrinsics, tilt, and satellite sampling scale are fixed for each altitude, these calibrated sizes can be reused across different target locations. Moreover, Google Earth's LookAt parameter keeps the camera focused on the same geographic target center across different headings and altitudes, ensuring consistent target-region correspondence across the 32 UAV views. The resulting annotations and coordinate transformations are manually checked for spatial consistency. We have clarified the camera configuration and bounding-box calibration procedure in Appendices A.3 (lines 752–756) and A.4 (lines 766–787) of the revised manuscript.

Q2: The practical assumptions behind gallery construction need further clarification. The main evaluation provides each query with 100 candidate tiles, including its paired tile. In practical use, how would the geographic coverage of the satellite gallery be determined before localizing the query? Would this require prior knowledge of the operating area or an approximate UAV location? Does the method assume that the target is always within the predefined gallery?

A2: Thank you for these questions. First, regarding gallery coverage, the satellite gallery is defined based on the UAV's planned operating area, following the area-of-interest formulation of VIGOR. For example, if a UAV is deployed in a particular urban region, satellite tiles covering that region can be prepared in advance and used for all UAV queries. Second, regarding prior location information, the 100-tile main setting is a controlled sampling evaluation (covering ∼78,km2\sim 78\\,\mathrm{km}^2), whereas in practical deployment our method requires knowledge of the general operating area, but does not require the approximate location of each UAV query. All queries search the same satellite gallery without query-specific location information. If approximate location information is available, it can optionally narrow the search. Third, regarding whether the target is within the gallery, our current evaluation assumes that the satellite gallery contains the correct tile for each UAV query. If the target lies outside the gallery coverage, the current method does not explicitly identify such cases.

To further clarify the distinction between controlled evaluation and practical gallery construction, we evaluate UniSAG on 13 fixed neighboring-city galleries (Appendix C.2, lines 911–943; 37,248 queries across 1,165 tiles) and on the complete 2,804-tile test gallery (Section 5.3, lines 472–479; Table 10, lines 498–505), without query-specific gallery filtering. UniSAG (full) achieves 24.55% macro R@1 (24.12% micro R@1) on the 13-city benchmark and 34.83% R@1 (14.65% uIoU) on the complete 2,804-tile gallery. We have clarified gallery construction and evaluation assumptions in Section 5 (lines 288–290) and Appendix C.2 (lines 912–916) of the revised manuscript.

Q3: Both the UAV and satellite images are generated using the same Google Earth rendering pipeline, which leaves uncertainty about how well the results generalize to real UAV imagery. Could the authors discuss whether this shared source reduces the difficulty of cross-view matching, and what implications it has for applying the method to real UAV imagery?

A3: When collecting UAV images in Google Earth, we enable the "3D Buildings" option to obtain textured 3D building models, whereas this option is disabled for satellite images. These different acquisition settings introduce clear visual and geometric differences between the two views. Moreover, WideUAV includes wide-area satellite coverage, off-center targets, and visually similar neighboring regions, making cross-view matching challenging despite the shared rendering environment.

To examine generalization to real UAV imagery, we conduct a zero-shot transfer experiment using 93 real UAV queries from SUES-200 at 150 m altitude, paired with corresponding wide-area Google Earth satellite tiles. UniSAG and eight external baselines are trained exclusively on WideUAV and evaluated on SUES-200 without fine-tuning, using the same model weights. Each target region uses a 330×330330\times330 pixel box, and headings are estimated from the paired images and quantized to 45∘45^\circ. The 93 non-overlapping satellite tiles are supplemented with negative tiles to form a 100-candidate gallery. On WideUAV, we report the 13-city macro R@1 (from Table 15, lines 927–943) alongside paired-tile oracle mIoU. All grounding baselines and UniSAG receive the same heading and altitude inputs.

ModelWideUAV R@1 (%)WideUAV mIoU (%)SUES-200 R@1 (%)SUES-200 mIoU (%)
CLIP11.45–6.45–
SigLIP29.02–5.38–
Sample4Geo (Retrieval)10.88–5.38–
DetGeo–4.78–6.00
TROGeo–4.31–3.17
Sample4Geo (Grounding)–6.71–9.56
TransGeo15.423.918.603.00
UnifyGeo8.603.264.305.50
UniSAG (full)24.5519.3630.1117.52

As shown above, UniSAG achieves 30.11% R@1 and 17.52% mIoU on real SUES-200 UAV queries, outperforming the strongest corresponding baselines by 21.51 and 7.96 percentage points, respectively. These results demonstrate that UniSAG retains its retrieval and grounding advantages on real UAV imagery without adaptation, rather than relying exclusively on Google Earth-generated UAV queries. We have clarified the acquisition settings in Appendix A.1 (lines 729–735) and added the zero-shot transfer experiment in Appendix C.3 (lines 959–1006; Table 17 at lines 981–994) of the revised manuscript.

Thank you for your detailed response. My concerns have been addressed, and I will maintain my current positive rating.

We sincerely thank you for the positive follow-up and are delighted that our response and revisions have fully addressed your concerns. We truly appreciate your recognition of the application value of our work, as well as your continued support and constructive feedback throughout the review process.

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

This manuscript introduced WideUAV, a rendered benchmark for UAV-to-satellite tile retrieval and within-tile target grounding, and proposes UniSAG to address both tasks jointly. UniSAG combines regional retrieval with heading- and altitude-conditioned grounding through shared visual representations, using retrieval-gated uIoU to summarize performance across the two components.

Critical strengths

Their benchmark addressed a relevant setting in which the target must be localized within a large satellite region, extending beyond target-centered retrieval. Tab. 4 well shows that joint training improves retrieval R1 from 62.03% to 65.75%, and the appendix provides useful analyses of geographic proximity, gallery difficulty, target geometry, and uncertainty across repeated runs.

Critical weaknesses

I can not find enough statements about the benefit or effect of the unified model in this manuscript. The separate-encoder variant in Tab. 6 also changed the loss balance and text inputs, so its lower performance cannot be attributed to encoder sharing alone. I think that the authors can and should use matched objectives and inputs, and compare with a pipeline combining independently trained UniSAG-R and UniSAG-G, to establish the benefit of the joint model.

Furthermore, the main comparison (evaluation metric uIoU) does not always measure the actual pipeline output. The main uIoU uses paired-tile oracle grounding even when a different positive tile is accepted, so it need not reflect the prediction on the retrieved tile. I think the Tab. 10 is the most proper evaluation protocol. Tab. 10’s exact-pair protocol should be used consistently in the main comparisons, or grounding should be evaluated on the retrieved tile using the corresponding target coordinates.

Lastly, for top-tier conference paper, this manuscript needs stronger comparison controls. The main comparisons use only 100 candidates per query from about 3 thousands test tiles, and Tab. 5’s grounding baselines do not receive the heading information used by UniSAG.

Questions for authors

I'm not an expert in this research area, so please kindly address all of the concerns in weakness section. The benchmark and the accompanying analyses are useful, but the current comparisons looks not sufficiently establish the benefit of the unified model under matched conditions. The interpretation of the main metric and the wide-area evaluation also need to be strengthened, so I lean toward weak Reject in its current form.

In short, please address all of the issues in weakness section. If all of the concerns were cleaned in rebuttal, I will raise this rate to acceptance level.

2 comments

We sincerely thank the reviewer for the detailed review and professional insights into our evaluation protocol and experimental controls. We agree that these controls and protocol checks are important. Accordingly, we have conducted additional experiments, reported in the responses below and in Appendix G (lines 1159–1239), to examine whether these choices affect our conclusions, while retaining the original evaluation protocol in the main text. The updated manuscript has been uploaded with all revisions highlighted in blue, and our point-by-point responses are given below.

Q1: The benefits of the unified model are insufficiently demonstrated, as the separate-encoder comparison changes loss balance and text inputs. Matched comparisons with separate encoders and independently trained retrieval–grounding pipelines are needed.

A1: Thank you for pointing out this confounding factor in the original separate-encoder comparison. We have evaluated the separate-encoder model (which uses separate backbones for retrieval and grounding under joint training) with the same loss balance as UniSAG (full). We also introduce a two-stage cascade combining independently trained UniSAG-R and UniSAG-G, where UniSAG-R first retrieves the top-1 satellite tile and UniSAG-G predicts the target box within it.

Retrieval modelGrounding modelR@1 (%)uIoU (%)uAcc@25 (%)uAcc@50 (%)uCDE (m)↓\downarrow
Matched losses and inputs
Separate encodersSeparate encoders54.9917.4825.52–229.11
UniSAG (full)UniSAG (full)65.7522.6732.73–213.61
Two-stage inference cascades with UniSAG-G
UniSAG-RUniSAG-G60.4222.8632.0227.56208.87
UniSAG (full)UniSAG-G65.7024.7534.6829.88210.09

Under matched objectives and inputs, UniSAG (full) improves R@1 from 54.99% to 65.75% and uIoU from 17.48% to 22.67% compared with separate encoders, demonstrating the benefit of shared feature learning. In the two-stage cascade comparison (Table 6 at lines 378–392 and Section 5.2 at lines 373–377, 426–431, and 456–457 of the revised manuscript), replacing UniSAG-R with UniSAG (full) for retrieval while keeping the same UniSAG-G grounder improves R@1 from 60.42% to 65.70%, uIoU from 22.86% to 24.75%, and uAcc@25 from 32.02% to 34.68%, with a comparable center-distance error (210.09 m vs. 208.87 m) and no additional inference overhead.

Q2: Inconsistent uIoU evaluation when retrieval accepts overlapping satellite tiles but grounding is evaluated on the paired tile, rather than the retrieved tile.

A2: Thank you for raising this important point. In our dataset, satellite tiles in the gallery overlap with one another, so the same target region often appears in multiple tiles. While some neighboring tiles contain the complete target region, others capture only part of it, leaving the target only partially visible near the tile boundary. We cannot count such partially captured regions as successful retrievals. At the same time, it is not practical to annotate target boxes across all gallery tiles, because evaluating grounding on an incomplete target region would be unfair compared with a complete one.

Although we could remove all other tiles that cover the target region from the gallery so that only the paired tile remains, doing so would lose the multi-ground-truth characteristic that naturally exists in both our dataset and practical applications. For this reason, our main evaluation retains the multi-ground-truth setting in retrieval (where any tile that fully covers the target region is accepted) and treats accepted retrieval as a prerequisite for end-to-end grounding, while evaluating grounding uniformly on the exact paired tile.

Following your suggestion, we also conducted new experiments on our main results in Appendix G (lines 1161–1177; Tables 23–25 at lines 1190–1229) under the exact-pair protocol, where other tiles covering the target region are excluded from the 100-tile subset and both retrieval and grounding are evaluated strictly on the paired tile.

MethodR@1 (%)uIoU (%)uAcc@25 (%)uCDE (m)
Exact-pair protocol, 100 sampled tiles
UnifyGeo28.751.562.42342.71
TransGeo46.613.575.61326.36
UniSAG (full)64.3822.2732.31213.96
UniSAG-R + UniSAG-G59.8122.8832.07206.44
UniSAG (full) + UniSAG-G64.3824.3734.21208.02

As shown above and in Appendix G (lines 1159–1239), the new experiments under the reviewer's suggested protocol confirm the conclusions of the main paper.

Q3: Concerns about the limited 100-candidate retrieval evaluation and the need for stronger comparisons on larger satellite galleries.

A3: We have extended retrieval to the complete 2,804-tile test gallery, as well as intermediate gallery sizes of 400 and 600 tiles, using UniSAG (full) across all 89,728 test queries.

Gallery settingR@1 (%)uIoU (%)
Gallery-size evaluation
100 sampled tiles63.5821.01
400 sampled tiles47.1516.93
600 sampled tiles43.0315.90
All 2,804 gallery tiles34.8314.65

As shown above, retrieval R@1 is 63.58% with 100 sampled tiles, 47.15% with 400, 43.03% with 600, and 34.83% with the complete gallery. Correspondingly, retrieval-gated uIoU is 21.01%, 16.93%, 15.90%, and 14.65%. These results highlight the increasing difficulty of distinguishing visually similar satellite regions as the search gallery expands. Notably, even the 100 sampled tiles cover approximately 78,km278\\,\mathrm{km}^2, comparable to the total satellite coverage of some existing benchmarks in Table 1 (lines 71–84). These results follow the main analysis in Section 5.3 (lines 472–479), Figure 5A (lines 441–451), and Table 10 (lines 498–505).

Q4: Regarding comparison controls, Table 5's grounding baselines do not receive the heading information used by UniSAG.

A4: We agree that this is an important control. As an additional comparison, we have provided all five external grounding baselines with the same geometric information used by UniSAG, including UAV heading and altitude, and retrained them under these settings. The geometric-input controls are described in Appendix F.3 (lines 1154–1157), and the additional baseline results are reported in Appendix G (lines 1175–1177; Table 24 at lines 1207–1216) and summarized below.

MethodmIoU (%)Acc@50 (%)CDE (m)
Sample4Geo11.879.84291.63
SMGeo10.536.33304.80
OCGNet8.827.36325.83
DetGeo8.084.97333.96
TROGeo7.705.84334.64
UniSAG (full)30.8436.77238.05
UniSAG-G (used in cascade)34.4044.07212.72

As shown above, UniSAG (full) achieves 30.84% mIoU, outperforming the strongest external baseline (Sample4Geo, 11.87%) by 18.97 percentage points, even after providing all baselines with the same geometric information. Notably, UniSAG-G without heading conditioning and heatmap supervision still achieves 21.28% mIoU in the ablation (Table 5, line 347), outperforming all five external baselines. These results demonstrate that UniSAG's grounding advantage cannot be attributed solely to the additional geometric inputs.

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

This paper extends conventional UAV-to-satellite image retrieval to retrieving the correct satellite tile and localizing the target within it. The authors introduce WideUAV benchmark, containing 3,505 satellite tiles and 112,160 UAV queries generated using Google Earth virtual cameras. They further propose UniSAG, a unified framework that combines region-level retrieval supervision, geometry-conditioned feature matching, and bounding-box prediction.

Critical strengths

1.The paper extends conventional cross-view image retrieval to a more complete localization setting that requires both satellite-tile retrieval and within-tile target grounding. It addresses an important limitation of existing target-centered retrieval benchmarks. 2. The experiments provides useful observations about the interaction between retrieval and grounding supervision.

Critical weaknesses
  1. Both UAV and satellite images are generated using the same Google Earth rendering environment rather than independently acquired UAV and satellite observations. It limits the benchmark's ability to represent real-world cross-platform differences. Moreover, the target bounding boxes are calculated from geographic coordinates and UAV altitude rather than manually annotated object boundaries. Without validation on real UAV–satellite imagery, it is difficult to assess the reported improvements.
  2. UniSAG largely combines existing components, including SigLIP2, LoRA, regional contrastive learning, FiLM conditioning, feature correlation, Transformer fusion, and YOLO-style box prediction. The paper does not sufficiently establish a fundamentally new methodological contribution beyond their integration.
Questions for authors

1.Since both UAV and satellite images are rendered using Google Earth, can the authors provide experiments on independently acquired real UAV and satellite imagery ? 2What is the key methodological novelty of UniSAG beyond integrating existing methods? Can the authors provide more comparisons to demonstrate the contributions of the proposed mechanisms? 3. The main retrieval experiments use 100 sampled candidates, while performance is substantially lower in the fixed neighboring-city galleries. How does UniSAG perform when retrieving directly from the entire 2,804-tile test gallery?

3 comments

We sincerely appreciate the reviewer's detailed comments and constructive suggestions on our dataset and method. The updated manuscript has been uploaded with all revisions highlighted in blue, and we provide point-by-point responses below.

Q1: The use of the same Google Earth rendering environment for both UAV and satellite images, together with geometrically derived bounding boxes, raises concerns about real-world applicability and motivates validation on real UAV–satellite imagery.

A1: Although both views are collected using Google Earth, their acquisition settings are different. Specifically, we enable "3D Buildings" when capturing UAV views to obtain textured 3D building models, while disabling this option for top-down satellite imagery. This introduces visual and geometric differences between the two views despite their shared platform.

To examine whether models trained on our dataset can handle real imagery, we conduct a zero-shot transfer experiment on SUES-200. We directly train UniSAG and eight external baselines on WideUAV and then evaluate them on SUES-200. In this transfer setting, the testing queries are real captured UAV images from SUES-200 while the satellite images are still rendered from Google Earth, so the two views no longer come from the same platform. Each target region uses a human-annotated center and an estimated UAV heading quantized to 45∘45^\circ. As far as we know, there is currently no public dataset with large-scale pairs of real UAV images and real wide-area satellite images (satellite images in UAV-VisLoc and AnyVisLoc are sourced from Google Earth, drone views in SkyReg are rendered from Google Earth, and OpenCVL only contains ground-view queries).

ModelWideUAV R@1 (%)WideUAV mIoU (%)SUES-200 R@1 (%)SUES-200 mIoU (%)
CLIP11.45–6.45–
SigLIP29.02–5.38–
Sample4Geo (Retrieval)10.88–5.38–
DetGeo–4.78–6.00
TROGeo–4.31–3.17
Sample4Geo (Grounding)–6.71–9.56
TransGeo15.423.918.603.00
UnifyGeo8.603.264.305.50
UniSAG (full)24.5519.3630.1117.52

On WideUAV, we evaluate on the 13 fixed city galleries (37,248 queries across 1,165 tiles), reporting macro-averaged R@1 and paired-tile oracle mIoU.

As shown above, UniSAG achieves 30.11% R@1 and 17.52% mIoU on real SUES-200 UAV queries, outperforming the strongest corresponding baselines by 21.51 and 7.96 percentage points, respectively. Compared with its results on WideUAV (24.55% macro R@1 and 19.36% oracle mIoU), UniSAG retains strong retrieval and target-region grounding performance when transferred to real UAV imagery without fine-tuning. These results provide evidence that UniSAG's advantages are not limited to UAV queries generated within Google Earth. We have clarified the acquisition settings in Appendix A.1 (lines 729–735) and detailed the experimental setup in Appendix C.3 (lines 959–1006; Table 17 at lines 981–994) of the revised manuscript.

Regarding bounding-box annotations, we clarify that the boxes represent target regions observed by UAV queries rather than precise boundaries of individual buildings, as some buildings extend beyond the entire UAV view and some target regions depict highways or ponds rather than specific buildings. Their centers are determined from geographic coordinates, while their extents are manually calibrated by matching corresponding reference points across four UAV views (0∘0^\circ, 90∘90^\circ, 180∘180^\circ, and 270∘270^\circ) and the satellite image. The resulting annotations and coordinate transformations are manually checked to ensure spatial consistency. We have clarified the annotation semantics and calibration procedure in Appendix A.4 (lines 766–787) of the revised manuscript.

Q2: Methodological novelty beyond integrating existing components and additional comparisons to validate the proposed mechanisms.

A2: We thank the reviewer for this comment. The contribution of UniSAG lies in how retrieval and grounding are coupled to address wide-area cross-view geo-grounding, rather than in introducing these components individually. Specifically, UniSAG couples target-aware regional retrieval with geometry-conditioned grounding through a shared visual encoder, using heading-aligned feature correlation and heatmap supervision for spatial localization. This design enables feature reuse without pairwise cross-view fusion, requiring only n+mn+m feature extractions for nn UAV queries and mm satellite tiles instead of n×mn\times m, while reducing redundant backbone parameters. More importantly, joint learning not only supports both tasks but also improves retrieval performance through grounding supervision.

We provide several comparisons to demonstrate these contributions. First, joint training improves R@1 from 62.03% with UniSAG-R to 65.75% with UniSAG (full) (Table 4, lines 324–335), demonstrating the benefit of grounding supervision. Second, the shared-encoder model achieves 65.75% R@1 and 22.67% uIoU, compared with 54.99% and 17.48% for the separate-encoder variant (Table 6, lines 378–392). Third, joint training outperforms sequential training, which achieves only 5.05% and 10.53% uIoU under the two training orders (Table 6). Furthermore, combining heading conditioning and heatmap supervision improves grounding mIoU from 21.28% to 34.40% in the grounding-only model (Table 5, lines 337–351) with negligible additional parameters or computation. These results support the benefits of coupling retrieval and grounding through shared representations. The main comparisons are reported in Sections 5.1 and 5.2 (lines 319–377, 426–431, and 456–468; see also feature-reuse complexity in Section 4, lines 241–245).

Q3: The main retrieval evaluation uses 100 sampled tiles, while performance drops substantially on fixed neighboring-city galleries. How does UniSAG perform on the complete 2,804-tile test gallery?

A3: Our main evaluation uses 100 sampled tiles, covering approximately 78,km278\\,\mathrm{km}^2, which is already comparable to the total satellite coverage of some existing benchmarks in Table 1 (lines 71–84). To examine how performance scales with gallery size, we further evaluate UniSAG (full) across all 89,728 test queries as the candidate pool expands from 100 to 400, 600, and all 2,804 test gallery tiles (Figure 5A at lines 441–451 and Table 10 at lines 498–505 of the revised manuscript).

Gallery settingR@1 (%)uIoU (%)
100 sampled tiles63.5821.01
400 sampled tiles47.1516.93
600 sampled tiles43.0315.90
All 2,804 gallery tiles34.8314.65

As shown in Figure 5A and Table 10 (lines 498–505; see Section 5.3, lines 472–479), R@1 decreases from 63.58% with 100 sampled tiles to 47.15% with 400 and 43.03% with 600, and reaches 34.83% when searching all 2,804 gallery tiles. Correspondingly, retrieval-gated uIoU decreases from 21.01% to 16.93%, 15.90%, and 14.65%. These results show that end-to-end geo-grounding remains effective on the complete 2,804-tile gallery while quantifying how visually similar distractors accumulate as the search area expands, consistent with the fixed-city gallery analysis in Appendix C.2 (lines 911–955; Table 15 at lines 927–943).

Thanks again for your time and constructive feedback on our work. As the author-reviewer discussion deadline is approaching, we would like to kindly check whether our response and revised manuscript have addressed your concerns, or if there are any remaining questions we can further clarify.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.