Pith. sign in

REVIEW 3 major objections 6 minor 45 references

A new benchmark and framework make UAV-to-satellite localization viable at steep off-nadir angles, lifting Recall@1 to 72.64%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A new benchmark and framework for matching sharply angled UAV photos to satellite maps, with an aggregation and group-learning method that outperforms prior work on most tested datasets.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Useful new synthetic benchmark for large off-nadir UAV-to-satellite matching, but the method's overclaimed 'consistent' superiority and the lack of real oblique validation temper the practical claims. the 3 major comments →

arxiv 2607.19951 v1 pith:BNG5USAN submitted 2026-07-22 cs.CV

OffNadirLoc: Benchmark and Framework for Challenging UAV-to-Satellite Geo-Localization under Large Off-Nadir Views

Qian Qiao , Wenye Liu , Ting Liu , Jiuhe Shu , Peng Wang This is my paper
classification cs.CV
keywords off-nadirUAV-to-satellite geo-localizationcross-view matchingbenchmarkstructure-aware weightingview-coherent learningmulti-view supervisionDINOv2
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that large off-nadir UAV-to-satellite geo-localization fails because existing methods aggregate features indiscriminately and supervise only pairwise image relations, and that fixing both with structure-aware aggregation and group-level supervision restores accuracy. To test this, it introduces OffNadirLoc, a benchmark of about 9,700 UAV images rendered at 70-85 degrees off-nadir with corresponding satellite patches across 44 regions. The proposed ONLoc framework weights local features by redundancy before clustering them into structure descriptors, and trains by treating all UAV views of a location with its satellite image as one semantic group. On the benchmark ONLoc reaches 72.64% Recall@1 with a DINOv2 encoder, surpassing the second-best method by 6.9 points, and it transfers zero-shot to four near-nadir datasets. If the benchmark's synthetic renderings faithfully represent real oblique UAV photos, the work shows a concrete recipe for making cross-view matching robust to extreme perspective.

Core claim

The central claim is that a combination of two mechanisms makes large off-nadir UAV-to-satellite matching tractable: a structure-aware contextual weighting (SACW) module that estimates per-token redundancy via a self-similarity matrix and down-weights repetitive regions during soft clustering, and a view-coherent learning strategy (VCLS) that treats all M UAV views of a location plus its satellite image as one semantic group, mining all intra-group positive pairs and inter-group negatives under a Multi-Similarity loss. With a DINOv2 encoder, this combined framework achieves state-of-the-art results on the new OffNadirLoc benchmark (72.64% R@1, 73.34% AP), and trained only on OffNadirLoc it t

What carries the argument

The load-bearing mechanism is SACW: local ViT features are L2-normalized, a pairwise similarity matrix S is computed, and a redundancy score per token is derived by summing sigmoid-activated similarities; a soft-assignment MLP then produces cluster memberships whose logits are divided by the redundancy score raised to a learnable exponent, so repetitive regions contribute less. The resulting K cluster descriptors are concatenated with the global token. The second mechanism, VCLS, replaces satellite-to-UAV pairwise contrast with set-level supervision: all M+1 images of a location form a positive set and all other locations form negatives, optimized with Multi-Similarity loss. Together, SACW d

Load-bearing premise

The benchmark's UAV images are Google Earth 3D renderings at fixed 70-85 degree off-nadir angles with simulated shadows and textures, and the paper assumes these are close enough to real oblique UAV imagery; if real photos contain lens distortion, motion blur, atmospheric haze, or physical occlusions that the renderings lack, the benchmark's validity and the method's claimed real-world generalization both collapse.

What would settle it

Collect real oblique UAV photos with known GPS locations over terrain comparable to OffNadirLoc regions, match them to the same satellite basemaps, and run ONLoc trained only on OffNadirLoc; if Recall@1 falls far below the 72.64% reported on synthetic queries, the central generalization claim is refuted. A cheaper check is computing a feature-based domain distance between OffNadirLoc renderings and real UAV photos; a large gap would predict the drop.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • On OffNadirLoc, ONLoc reaches 72.64% R@1 and 73.34% AP with a DINOv2 backbone, 6.9 points ahead of the second-best method in R@1.
  • Trained solely on OffNadirLoc, the model transfers zero-shot to University-1652 (79.87% R@1), SUES-200 (91.38%), DenseUAV (18.28%), and GTA-UAV (56.82%), beating prior zero-shot methods in most cases.
  • Ablations show SACW alone adds +8.42% R@1 over the DINOv2 baseline, VCLS alone adds +4.29%, and combining them yields +10.35% R@1 and +7.71% AP.
  • The method stays competitive when trained and tested on standard near-nadir benchmarks, especially DenseUAV (88.72% R@1) and GTA-UAV (65.29% R@1).
  • Performance degrades smoothly across all off-nadir angles from 70 to 85 degrees, whereas prior methods drop sharply with tilt.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If confirmed on real oblique UAV photos, redundancy suppression plus group-level supervision could generalize beyond UAV-to-satellite retrieval to any cross-view task where multiple images of one scene exist, such as ground-to-aerial place recognition.
  • The benchmark's synthetic rendering pipeline could be extended to include lens distortion, motion blur, atmospheric effects, and physical occlusions; a test with real collected oblique imagery would verify whether the 70-85 degree range captures the difficulty of actual deployments.
  • The group-level supervision idea implies that when several drones view the same site, one UAV's image might be localized not only against satellite imagery but against the aggregate of other drones' views, potentially improving robustness without additional satellite data.
  • The large zero-shot gap on DenseUAV (18.28% R@1) suggests that dense-patch evaluation protocols reward near-perfect alignment, and models trained on sparse off-nadir renderings may miss fine-grained spatial alignment; a protocol with softer spatial tolerance might change the ranking.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces OffNadirLoc, a UAV-to-satellite geo-localization benchmark rendered from Google Earth 3D imagery at extreme off-nadir angles (70°–85°), with a partial-matching evaluation protocol and a region-level train/test split. To address the benchmark, the authors propose ONLoc, which combines a structure-aware contextual weighting (SACW) module that clusters local features with redundancy suppression, and a view-coherent learning strategy (VCLS) that treats all UAV views of a location plus the satellite image as a semantic group under a multi-similarity loss. Experiments report state-of-the-art results on OffNadirLoc with both ConvNeXt and DINOv2 backbones, and zero-shot and supervised results on four near-nadir benchmarks (University-1652, SUES-200, DenseUAV, GTA-UAV). The paper claims consistent out-performance of existing methods and strong zero-shot generalization.

Significance. If the benchmark is a faithful proxy for real off-nadir UAV imagery, OffNadirLoc would fill a clear gap in cross-view geo-localization evaluation, and the ONLoc framework's combination of redundancy-aware aggregation and set-level supervision is a reasonable and potentially transferable design. The region-level split and partial-overlap evaluation are sensible methodological choices, and the controlled comparison under two backbones is a strength. However, the central empirical claims are weakened by the fact that the benchmark is generated from the same Google Earth 3D engine for queries and gallery, so positive pairs share basemap texture and rendering state; no real oblique UAV data validate the 70–85° regime. In addition, the paper's own Table 3 contradicts the 'consistently outperforms' wording in several rows, and all results are single-run without variance estimates. The benchmark may still be useful as a synthetic stress test, but the practical-capability claim requires external validation or substantial softening.

major comments (3)
  1. [Abstract, §5.2–5.3, Table 3] The claim that ONLoc 'consistently outperforms' SOTA and shows 'strong zero-shot generalization' is contradicted by the paper's own Table 3. In zero-shot DenseUAV, ONLoc achieves R@1=18.28% and R@5=49.29%, below ConGEO (22.31%, 55.73%) and, for R@5, below MEAN (49.21%). In supervised SUES-200, ONLoc R@1=97.99% and R@5=99.40% are below MEAN (98.56%, 99.93%). These are not isolated edge cases: several margins elsewhere are under 1%. All numbers are single-run with no error bars, so ranking stability is unknown. The abstract and §5.3 should be revised to report dataset-specific outcomes rather than a universal claim.
  2. [§3.1, §5.3] The benchmark's validity for real off-nadir UAV localization is not established. Both UAV queries and satellite gallery are produced from Google Earth's 3D engine for the same regions, so every positive pair shares the same basemap texture, building models, and lighting state. This permits a model to succeed by matching low-level perspective-warped texture rather than by learning viewpoint-invariant structure. Zero-shot tests on University-1652 are weak evidence because that dataset is itself Google Earth oblique imagery, and the other zero-shot datasets are near-nadir. Please validate on real oblique UAV imagery at 70–85°, or explicitly scope the claims to simulated off-nadir imagery and discuss the domain gap.
  3. [§5.5, Table 5] The hyperparameter K (number of SACW clusters) is selected by an ablation on OffNadirLoc reported in Table 5. If this ablation uses the same test split on which final results are reported, the test performance is optimistically biased. The paper must state whether a separate validation split or cross-validation was used to select K. This is load-bearing because K=64 is not a prior design choice but is tuned on the benchmark that also produces the headline numbers.
minor comments (6)
  1. [Throughout] The text repeatedly contains 'UA V' with an erroneous space; please fix to 'UAV' throughout.
  2. [§5.2] The sentence 'surpassing the second-best approach by 6.9% in Recall@1 and 4.75% in AP' should specify that the margins refer to the DINOv2 backbone. Under ConvNeXt the margins are smaller (e.g., R@1 gain is 2.95%).
  3. [§3.1] Please report the number of training/test regions, the number of locations per region, and the distribution of off-nadir angles. Currently only totals (9,736 UAV, 1,657 satellite) are given, which is insufficient for assessing dataset difficulty and split balance.
  4. [§5.1] The input resolution is 1000×600, which is not divisible by the DINOv2 patch size of 14. State the preprocessing (resize, padding, or interpolation) used for both training and evaluation.
  5. [Table 1] The notation 'Perfect×' and 'Partial✓' under 'Evaluation Protocol' is cryptic. Define these terms in the caption or text.
  6. [§4.2, Eq. (7)] Equation (7) leaves ℓ_p and ℓ_n unspecified. Provide the exact forms or a precise reference to the Multi-Similarity loss with the adaptation to group-wise positives/negatives so the objective is reproducible.

Circularity Check

0 steps flagged

No significant circularity; ONLoc's gains are empirical and externally benchmarked. The two suspect choices (same-source synthetic testbed and K=64 test-set tuning) are dataset-validity/reporting concerns, not by-construction reductions.

full rationale

ONLoc's representation (Eqs. 2–6) is a weighted aggregation of DINOv2 features with learned clusters; the redundancy score and cluster assignment depend only on input features, not on benchmark labels or retrieval metrics. The view-coherent loss (Eq. 7) is a Multi-Similarity objective over dataset-defined groups, i.e. standard supervised training, not a quantity derived from the reported results. The empirical claims are checked on OffNadirLoc and on four external benchmarks (University-1652, SUES-200, DenseUAV, GTA-UAV) against six cited baselines; there is no self-citation chain or imported uniqueness theorem. The two flagged concerns — (a) both UAV queries and satellite gallery come from Google Earth 3D rendering of the same scene, and (b) the cluster count K=64 is swept on the OffNadirLoc benchmark (Table 5) and then reused in the headline Table 2 — are genuine dataset-validity and test-set-selection concerns, but they are not circular derivations: no equation in the paper defines the benchmark or the method's output in terms of the reported performance. Under the required reduction standard, these do not qualify as circular steps, so the score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central empirical claim sits on the benchmark construction, the redundancy heuristic, and the group-supervision assumption; none of these are derived from first principles.

free parameters (5)
  • Number of SACW clusters K = 64
    Chosen by ablation on the OffNadirLoc evaluation set (Table 5); peak at 64, no separate validation split, so this is tuned on the benchmark it is evaluated on.
  • Learnable redundancy scalars a, b in Eq. (3)
    Trained parameters of the redundancy score; they shape the weighting mechanism but are learned, not fitted to final metrics.
  • Learnable redundancy suppression exponent p in Eq. (4)
    Trained parameter controlling how strongly redundant features are suppressed.
  • Off-nadir angles {70,75,80,82,85} degrees = set of angles
    Chosen by the authors for dataset design; defines the benchmark's difficulty envelope.
  • IoU acceptance threshold = 0.14
    Borrowed from VIGOR protocol; determines what counts as a correct retrieval and therefore all reported numbers.
axioms (4)
  • domain assumption Google Earth 3D renderings at 70–85° off-nadir with simulated shadows/textures are a faithful proxy for real off-nadir UAV imagery.
    Load-bearing for OffNadirLoc's validity; introduced in §3.1. If renderings miss real lens distortion, motion blur, atmospheric effects, or physical occlusions, conclusions about real-world deployment do not transfer.
  • domain assumption Self-similarity of local features indicates geometric redundancy, so down-weighting mutually similar patches improves cross-view matching.
    Core heuristic of SACW (§4.1, Eq. 2–4). No proof that self-similarity correlates with geometric unreliability under off-nadir distortion.
  • domain assumption Grouping all UAV views plus satellite of a location as positives and using Multi-Similarity loss learns viewpoint-invariant embeddings.
    Central to VCLS (§4.2); relies on the multi-view grouping being semantically coherent and on the loss's mining behavior.
  • domain assumption DINOv2 pretrained features transfer to synthetic Google Earth-rendered UAV imagery.
    All methods use DINOv2/ConvNeXt; if synthetic domain differs from pretraining, all results shift. §5.1.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of OffNadirLoc: Benchmark and Framework for Challenging UAV-to-Satellite Geo-Localization under Large Off-Nadir Views." pith.science (2026). https://pith.science/paper/BNG5USAN

@misc{pith2026260719951,
  author       = {Pith},
  title        = {Pith review of: OffNadirLoc: Benchmark and Framework for Challenging UAV-to-Satellite Geo-Localization under Large Off-Nadir Views},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BNG5USAN}},
  note         = {Machine review of arXiv:2607.19951}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Cross-view geo-localization between UAV and satellite imagery remains a fundamental yet highly challenging task, especially under large off-nadir views where drastic perspective distortions, occlusions, and appearance gaps occur. Existing benchmarks and methods primarily focus on near-nadir scenarios and often overlook the importance of structural scene understanding and intra-domain relational constraints, limiting their performance in real-world deployments. In this work, we introduce OffNadirLoc, a new benchmark for large off-nadir UAV-to-satellite geo-localization. To tackle the unique challenges posed by off-nadir perspectives, we further propose ONLoc, a framework that incorporates a structure-aware contextual weighting mechanism to dynamically emphasize reliable local features while suppressing ambiguous or repetitive regions. Additionally, we design a view-coherent learning strategy, which treats one satellite image and the corresponding UAV images from multiple views as a cohesive semantic group. This set-level supervision enables the model to learn viewpoint-invariant and discriminative features, making it more effective at capturing multi-view consistency than conventional pairwise contrastive learning. Extensive experiments on the OffNadirLoc benchmark and four near-nadir datasets demonstrate that our method consistently outperforms state-of-the-art approaches while exhibiting strong zero-shot generalization to unseen datasets without additional training. The code will be released at https://montalario.github.io/offnadirloc/.

Figures

Figures reproduced from arXiv: 2607.19951 by Jiuhe Shu, Peng Wang, Qian Qiao, Ting Liu, Wenye Liu.

Figure 1
Figure 1. Figure 1: OffNadirLoc benchmark overview. UAV images captured under large off-nadir angles (70°, 75°, 80°, 82°, 85°) exhibit severe [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the ONLoc pipeline. (a) Structure-Aware Contextual Weighting: Local and global features are extracted and [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The top row presents similarity distribution visualizations for different methods, while the bottom row shows feature embeddings [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance on the OffNadirLoc benchmark under large [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 7 linked inside Pith

  1. [1]

    Netvlad: Cnn architecture for weakly supervised place recognition

    Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pa- jdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5297–5307, 2016. 3

    2016

  2. [2]

    Re- thinking visual geo-localization for large-scale applications

    Gabriele Berton, Carlo Masone, and Barbara Caputo. Re- thinking visual geo-localization for large-scale applications. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4878–4888, 2022. 1

    2022

  3. [3]

    Deep visual geo-localization benchmark

    Gabriele Berton, Riccardo Mereu, Gabriele Trivigno, Carlo Masone, Gabriela Csurka, Torsten Sattler, and Barbara Ca- puto. Deep visual geo-localization benchmark. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5396–5407, 2022. 1

    2022

  4. [4]

    Multi- level embedding and alignment network with consistency and invariance learning for cross-view geo-localization

    Zhongwei Chen, Zhao-Xu Yang, and Hai-Jun Rong. Multi- level embedding and alignment network with consistency and invariance learning for cross-view geo-localization. IEEE Transactions on Geoscience and Remote Sensing,

  5. [5]

    Crowd-sourced pic- tures geo-localization method based on street view images and 3d reconstruction.ISPRS journal of photogrammetry and remote sensing, 141:72–85, 2018

    Liang Cheng, Yi Yuan, Nan Xia, Song Chen, Yanming Chen, Kang Yang, Lei Ma, and Manchun Li. Crowd-sourced pic- tures geo-localization method based on street view images and 3d reconstruction.ISPRS journal of photogrammetry and remote sensing, 141:72–85, 2018. 2

    2018

  6. [6]

    Ming Dai, Jianhong Hu, Jiedong Zhuang, and Enhui Zheng. A transformer-based feature segmentation and region align- ment method for uav-view geo-localization.IEEE Transac- tions on Circuits and Systems for Video Technology, 32(7): 4376–4389, 2021. 3

    2021

  7. [7]

    Vision-based uav self- positioning in low-altitude urban environments.IEEE Trans- actions on Image Processing, 33:493–508, 2023

    Ming Dai, Enhui Zheng, Zhenhua Feng, Lei Qi, Jiedong Zhuang, and Wankou Yang. Vision-based uav self- positioning in low-altitude urban environments.IEEE Trans- actions on Image Processing, 33:493–508, 2023. 2, 6, 7

    2023

  8. [8]

    Sam- ple4geo: Hard negative sampling for cross-view geo- localisation

    Fabian Deuser, Konrad Habel, and Norbert Oswald. Sam- ple4geo: Hard negative sampling for cross-view geo- localisation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 16847–16856, 2023. 1, 3, 6, 7

    2023

  9. [9]

    A practical cross-view image matching method between uav and satellite for uav-based geo-localization.Remote Sensing, 13(1):47, 2020

    Lirong Ding, Ji Zhou, Lingxuan Meng, and Zhiyong Long. A practical cross-view image matching method between uav and satellite for uav-based geo-localization.Remote Sensing, 13(1):47, 2020. 3

    2020

  10. [10]

    An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 4

    Pith/arXiv arXiv 2010

  11. [11]

    Cross-view geo-localization: a survey

    Abhilash Durgam, Sidike Paheding, Vikas Dhiman, and Vi- jay Devabhaktuni. Cross-view geo-localization: a survey. IEEE Access, 2024. 1

    2024

  12. [12]

    Cvm-net: Cross-view matching network for image- based ground-to-aerial geo-localization

    Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image- based ground-to-aerial geo-localization. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258–7267, 2018. 3

    2018

  13. [13]

    Learning where to learn in cross-view self-supervised learning

    Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Learning where to learn in cross-view self-supervised learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14451–14460, 2022. 3

    2022

  14. [14]

    Optimal transport ag- gregation for visual place recognition

    Sergio Izquierdo and Javier Civera. Optimal transport ag- gregation for visual place recognition. InProceedings of the ieee/cvf conference on computer vision and pattern recogni- tion, pages 17658–17668, 2024. 4

    2024

  15. [15]

    Game4loc: A uav geo-localization benchmark from game data

    Yuxiang Ji, Boyong He, Zhuoyue Tan, and Liaoni Wu. Game4loc: A uav geo-localization benchmark from game data. InProceedings of the AAAI Conference on Artificial Intelligence, pages 3913–3921, 2025. 1, 2, 6, 7

    2025

  16. [16]

    Vlad-buff: burst-aware fast feature aggre- gation for visual place recognition

    Ahmad Khaliq, Ming Xu, Stephen Hausler, Michael Milford, and Sourav Garg. Vlad-buff: burst-aware fast feature aggre- gation for visual place recognition. InEuropean Conference on Computer Vision, pages 447–466. Springer, 2024. 4

    2024

  17. [17]

    Supervised contrastive learning.Advances in neural information processing systems, 33:18661–18673,

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning.Advances in neural information processing systems, 33:18661–18673,

  18. [18]

    Unleashing un- labeled data: A paradigm for cross-view geo-localization

    Guopeng Li, Ming Qian, and Gui-Song Xia. Unleashing un- labeled data: A paradigm for cross-view geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 16719–16729, 2024. 1

    2024

  19. [19]

    Robust cross- view geo-localization via content-viewpoint disentangle- ment.arXiv preprint arXiv:2505.11822, 2025

    Ke Li, Di Wang, Xiaowei Wang, Zhihong Wu, Yiming Zhang, Yifeng Wang, and Quan Wang. Robust cross- view geo-localization via content-viewpoint disentangle- ment.arXiv preprint arXiv:2505.11822, 2025. 1

    arXiv 2025

  20. [20]

    Cross-view image geolocalization

    Tsung-Yi Lin, Serge Belongie, and James Hays. Cross-view image geolocalization. InProceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pages 891–898, 2013. 2

    2013

  21. [21]

    Lending orientation to neural networks for cross-view geo-localization

    Liu Liu and Hongdong Li. Lending orientation to neural networks for cross-view geo-localization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5624–5633, 2019. 2

    2019

  22. [22]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 5

    Pith/arXiv arXiv 2017

  23. [23]

    Congeo: Robust cross-view geo-localization across ground view vari- ations.arXiv preprint arXiv:2403.13965, 2024

    Li Mi, Chang Xu, Javiera Castillo-Navarro, Syrielle Montar- iol, Wen Yang, Antoine Bosselut, and Devis Tuia. Congeo: Robust cross-view geo-localization across ground view vari- ations.arXiv preprint arXiv:2403.13965, 2024. 3, 6, 7

    Pith/arXiv arXiv 2024

  24. [24]

    Repre- sentation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Repre- sentation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018. 3

    Pith/arXiv arXiv 2018

  25. [25]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5

    Pith/arXiv arXiv 2023

  26. [26]

    Google map aided visual navigation for uavs in gps-denied environment

    Mo Shan, Fei Wang, Feng Lin, Zhi Gao, Ya Z Tang, and Ben M Chen. Google map aided visual navigation for uavs in gps-denied environment. In2015 IEEE international confer- ence on robotics and biomimetics (ROBIO), pages 114–119. IEEE, 2015. 2 9

    2015

  27. [27]

    Spatial- aware feature aggregation for cross-view image based geo- localization.Advances in Neural Information Processing Systems, 32, 2019

    Yujiao Shi, Liu Liu, Xin Yu, and Hongdong Li. Spatial- aware feature aggregation for cross-view image based geo- localization.Advances in Neural Information Processing Systems, 32, 2019. 3

    2019

  28. [28]

    Where am i looking at? joint location and orientation es- timation by cross-view matching

    Yujiao Shi, Xin Yu, Dylan Campbell, and Hongdong Li. Where am i looking at? joint location and orientation es- timation by cross-view matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4064–4072, 2020. 3

    2020

  29. [29]

    Uav-satellite view synthesis for cross-view geo- localization.IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4804–4815, 2021

    Xiaoyang Tian, Jie Shao, Deqiang Ouyang, and Heng Tao Shen. Uav-satellite view synthesis for cross-view geo- localization.IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4804–4815, 2021. 1

    2021

  30. [30]

    Coming down to earth: Satellite-to-street view synthesis for geo-localization

    Aysim Toker, Qunjie Zhou, Maxim Maximov, and Laura Leal-Taix´e. Coming down to earth: Satellite-to-street view synthesis for geo-localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488–6497, 2021. 2

    2021

  31. [31]

    Multi-similarity loss with general pair weighting for deep metric learning

    Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R Scott. Multi-similarity loss with general pair weighting for deep metric learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5022–5030, 2019. 5

    2019

  32. [32]

    Image and object geo-localization.International Journal of Computer Vision, 132(4):1350–1392, 2024

    Daniel Wilson, Xiaohan Zhang, Waqas Sultani, and Safwan Wshah. Image and object geo-localization.International Journal of Computer Vision, 132(4):1350–1392, 2024. 1

    2024

  33. [33]

    Wide-area image geolocalization with aerial reference im- agery

    Scott Workman, Richard Souvenir, and Nathan Jacobs. Wide-area image geolocalization with aerial reference im- agery. InProceedings of the IEEE International Conference on Computer Vision, pages 3961–3969, 2015. 2

    2015

  34. [34]

    Camp: A cross-view geo-localization method using contrastive attributes mining and position-aware partitioning.IEEE Transactions on Geo- science and Remote Sensing, 2024

    Qiong Wu, Yi Wan, Zhi Zheng, Yongjun Zhang, Guang- shuai Wang, and Zhenyang Zhao. Camp: A cross-view geo-localization method using contrastive attributes mining and position-aware partitioning.IEEE Transactions on Geo- science and Remote Sensing, 2024. 1, 3, 6, 7

    2024

  35. [35]

    Enhancing cross-view geo-localization with do- main alignment and scene consistency.IEEE Transactions on Circuits and Systems for Video Technology, 2024

    Panwang Xia, Yi Wan, Zhi Zheng, Yongjun Zhang, and Ji- wei Deng. Enhancing cross-view geo-localization with do- main alignment and scene consistency.IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 3, 6, 7

    2024

  36. [36]

    Enhancing cross-view geo-localization with do- main alignment and scene consistency.IEEE Transactions on Circuits and Systems for Video Technology, 2024

    Panwang Xia, Yi Wan, Zhi Zheng, Yongjun Zhang, and Ji- wei Deng. Enhancing cross-view geo-localization with do- main alignment and scene consistency.IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1

    2024

  37. [37]

    Uav-visloc: A large- scale dataset for uav visual localization.arXiv preprint arXiv:2405.11936, 2024

    Wenjia Xu, Yaxuan Yao, Jiaqi Cao, Zhiwei Wei, Chunbo Liu, Jiuniu Wang, and Mugen Peng. Uav-visloc: A large- scale dataset for uav visual localization.arXiv preprint arXiv:2405.11936, 2024. 2

    Pith/arXiv arXiv 2024

  38. [38]

    Cross-view geo-localization with layer-to-layer transformer.Advances in Neural Information Processing Systems, 34:29009–29020,

    Hongji Yang, Xiufan Lu, and Yingying Zhu. Cross-view geo-localization with layer-to-layer transformer.Advances in Neural Information Processing Systems, 34:29009–29020,

  39. [39]

    Aligning geometric spatial layout in cross-view geo-localization via feature re- combination

    Qingwang Zhang and Yingying Zhu. Aligning geometric spatial layout in cross-view geo-localization via feature re- combination. InProceedings of the AAAI Conference on Ar- tificial Intelligence, pages 7251–7259, 2024. 3

    2024

  40. [40]

    Cross-view geo-localization via learning disentangled geometric layout correspondence

    Xiaohan Zhang, Xingyu Li, Waqas Sultani, Yi Zhou, and Safwan Wshah. Cross-view geo-localization via learning disentangled geometric layout correspondence. InProceed- ings of the AAAI conference on artificial intelligence, pages 3480–3488, 2023. 6

    2023

  41. [41]

    University- 1652: A multi-view multi-source benchmark for drone- based geo-localization

    Zhedong Zheng, Yunchao Wei, and Yi Yang. University- 1652: A multi-view multi-source benchmark for drone- based geo-localization. InProceedings of the 28th ACM international conference on Multimedia, pages 1395–1403,

  42. [42]

    Sues-200: A multi-height multi- scene cross-view image benchmark across drone and satel- lite.IEEE Transactions on Circuits and Systems for Video Technology, 33(9):4825–4839, 2023

    Runzhe Zhu, Ling Yin, Mingze Yang, Fei Wu, Yuncheng Yang, and Wenbo Hu. Sues-200: A multi-height multi- scene cross-view image benchmark across drone and satel- lite.IEEE Transactions on Circuits and Systems for Video Technology, 33(9):4825–4839, 2023. 2, 6, 7

    2023

  43. [43]

    Vigor: Cross- view image geo-localization beyond one-to-one retrieval

    Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross- view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 3640–3649, 2021. 6

    2021

  44. [44]

    Transgeo: Trans- former is all you need for cross-view image geo-localization

    Sijie Zhu, Mubarak Shah, and Chen Chen. Transgeo: Trans- former is all you need for cross-view image geo-localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1162–1171, 2022. 3

    2022

  45. [45]

    Simple, effective and general: A new back- bone for cross-view image geo-localization.arXiv preprint arXiv:2302.01572, 2023

    Yingying Zhu, Hongji Yang, Yuxin Lu, and Qiang Huang. Simple, effective and general: A new back- bone for cross-view image geo-localization.arXiv preprint arXiv:2302.01572, 2023. 1 10

    Pith/arXiv arXiv 2023

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.