Pith. sign in

REVIEW 5 major objections 5 minor 31 references

A Unified Benchmark and Modality-Adaptive Network for Day-and-Night Drone-View Geo-Localization

T0 review · 5 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read This paper claims drone-view geo-localization can work for both daytime visible and nighttime infrared queries using one benchmark, IRCHN, and one network, MASTR-Net, that together bridge the modality gap and the viewpoint gap.

desk verdict IRCHN is a genuinely useful benchmark, but the SOTA claims for MASTR-Net rest on uncontrolled baselines and single-run numbers; the dataset needs to be released. read the letter →

arxiv 2607.25778 v1 pith:XBJRN4ZH submitted 2026-07-28 cs.CV

Songtianhao Xu , Zhongwei Chen , Zhao-Xu Yang , Weifeng Wang This is my paper
classification cs.CV
keywords drone-viewgeo-localizationvisible-infraredmatchingday-nightlocalizationcross-viewretrievalstate-spacemodeloptimaltransportbenchmarkmodalityadaptation
verification ladder T0 review T1 audit T2 compute T3 formal

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 sets out to make drone-view geo-localization work for both daytime visible images and nighttime infrared images within a single framework. To that end it introduces IRCHN, a benchmark of 8,820 locations in Hainan, China, where each location provides one visible drone image, one infrared drone image, and one satellite reference image aligned by geographic identity rather than pixel correspondence. It also proposes MASTR-Net, which combines modality-adaptive feature enhancement, bidirectional state-space relation modeling, and a soft optimal transport alignment loss to reduce both the visible-infrared appearance gap and the drone-to-satellite viewpoint gap. The paper reports that MASTR-Net outperforms existing methods on IRCHN in both visible and infrared retrieval, with the largest gains in the infrared setting, and stays competitive on two existing infrared benchmarks. A sympathetic reader would care because the benchmark enables, for the first time, evaluation of day-and-night localization on the same geographic locations under a consistent protocol.

What carries the argument

Three mechanisms carry the argument. Modality-Adaptive Feature Enhancement recalibrates channel responses and adds a bounded, input-dependent modulation residual to the spatial feature map while leaving the global descriptor untouched. Bidirectional Selective State-Space Relation Mixer flattens the enhanced feature map into tokens, runs a retention-based state-space update in forward and backward orders with an input-dependent gate and a learnable base-retention vector, fuses the two scans, and aggregates tokens using the global descriptor as a query. Soft Optimal Transport Relation Alignment converts the global similarity matrix into a balanced soft target via alternating row-column normali

What would settle it

Take a random sample of IRCHN triplets and have a human or a geometric check (satellite image footprint versus drone image footprint) verify that the visible and infrared drone images actually depict the same location as the satellite reference; if a meaningful fraction of triplets do not overlap in ground area, the benchmark's positive pairs are corrupted. Alternatively, run the same retrieval protocol with a model that uses only coarse scene-level features (for example, image-level histograms) and check whether it still achieves high Recall@1, which would indicate the benchmark can be solved

Watch

Extended reading notes

Core claim

The central claim is that a single retrieval model can localize a drone image against satellite references whether the query is a daytime visible photo or a nighttime infrared frame, provided the model explicitly (1) adapts its spatial features to the input modality while preserving the global descriptor, (2) captures long-range structural relations in both forward and backward scans under global-semantic guidance, and (3) aligns the relation-level similarity distribution to a balanced soft target derived from the global similarity matrix via alternating row-column normalization. The paper reports that this combination, named MASTR-Net, achieves the best Recall@1 and average precision on the

Load-bearing premise

The integrity of IRCHN rests on using GPS coordinates to declare that a visible drone image, an infrared drone image, and a satellite image show the same place; if those coordinates are inaccurate or the three images do not actually cover the same ground area, every positive pair is mislabeled and all reported retrieval numbers are unreliable.

Editorial extensions

If this is right

  • If the benchmark is sound, DVGL researchers gain a standard protocol to measure day/night performance without dataset-transfer confounds, since visible and infrared queries come from the same locations.
  • The method's larger infrared gains suggest that explicitly modulating features by modality and aligning relation structure transfers across illumination conditions; this recipe is testable on other multimodal retrieval tasks.
  • A single MASTR-Net model handles both visible and infrared queries, so a deployed system could switch sensing modalities at night without retraining.
  • Because IRCHN includes per-scene splits, future methods can be analyzed for failure modes in repetitive farmland, homogeneous water, dense forest, and cluttered urban scenes.
  • The SOTRA loss, as a soft-target distillation from global to relation branch, could be reused in other cross-view or cross-modal retrieval settings regardless of the backbone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 'aligned by geographic identity' labeling means positive pairs tolerate large viewpoint and scale differences, but it also opens the door to shortcut learning: a model might solve IRCHN by exploiting scene-type or region-specific biases rather than true geo-localization. A testable extension is to perform retrieval on held-out regions or with satellite references shuffled across scene categori
  • Since all IRCHN imagery comes from a single airborne platform and one geographic region (Hainan), the benchmark's diversity is real but regional; multi-region expansion would be a natural stress test for both the benchmark and the method.
  • The paper does not report failure cases by time of acquisition or thermal crossover (for example, water and vegetation swapping appearance); specifically probing such cases would clarify where the modality gap remains.
  • A direct comparison with a simpler fusion of visible and infrared features (such as concatenation or late fusion) would help isolate whether the state-space relation mixer is the key ingredient or whether any strong relation model suffices.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper introduces IRCHN, a drone-view geo-localization (DVGL) benchmark containing 26,460 images from 8,820 geographic locations, each with a visible drone image, an infrared drone image, and a corresponding satellite image. The authors argue that this is the first benchmark supporting visible and infrared queries under a unified protocol with consistent geographic content. They also propose MASTR-Net, which combines a shared ConvNeXt-Base encoder, a Modality-Adaptive Feature Enhancement (MAFE) module, a Bidirectional Selective State-Space Relation Mixer (BSSRM), and a Soft Optimal Transport Relation Alignment loss (SOTRA). The central claim is that MASTR-Net consistently outperforms state-of-the-art methods on IRCHN under both visible and infrared settings, and remains competitive on IR-VL328 and CVGL-RGBT.

Significance. If the empirical claims hold, the paper makes two useful contributions. First, IRCHN addresses a real gap: existing DVGL benchmarks either contain only visible imagery or provide infrared images without spatially aligned visible counterparts. A benchmark with aligned visible, infrared, and satellite images at the same locations is valuable for studying day/night and cross-modal geo-localization. Second, MASTR-Net is a nontrivial architecture that combines state-space relation modeling, optimal-transport-based soft targets, and modality-adaptive enhancement; the SOTRA loss is a form of self-distillation that is internally coherent, not circular. The code release is a positive step, although the dataset itself is not released. However, the empirical validation of the SOTA claim is weakened by uncontrolled baseline comparisons, lack of repeated runs, possible test-set tuning, and an unverified data-alignment premise.

major comments (5)
  1. [Table IV / §V-B] The central SOTA claim rests on Table IV, but the baselines' training configurations are not reported. §V-B gives details only for MASTR-Net: ConvNeXt-Base, 384×384 input, 20 epochs, AdamW, batch 128. If Sample4Geo, MEAN, MFRGN, CAMP, DAC, or SURFNet were run with smaller backbones, lower resolutions, or shorter schedules, the reported gains (e.g., 1.63 points VIS avg R@1, 2.61 points IR avg R@1) could be due to model capacity rather than the proposed modules. Please report the backbone, input resolution, schedule, and loss settings for every baseline, or rerun them under matched settings.
  2. [Tables IV and IX] All results appear to be single runs; no error bars, standard deviations, or multiple seeds are reported. For retrieval tasks, differences of 1–3 percentage points in R@1 can lie within seed-level noise, especially with batch size 128 and only 882 query locations. The claim of 'consistent improvements' across directions and modalities needs support from at least 3–5 seeds with mean±std, or a paired significance test on the per-query results.
  3. [§III-B / §III-C] The benchmark's integrity depends on the assumption that geographic coordinate identity correctly associates visible, infrared, and satellite images. §III-B states that images are 'aligned by geographic identity rather than by pixel-level correspondence,' but no quantitative validation is provided: no check of coordinate accuracy, no measure of actual ground-area overlap, and no examples showing challenging near-duplicate locations. Since IRCHN is not released, readers cannot verify this premise. Please add alignment statistics (e.g., estimated overlap, per-scene coordinate error) and release at least a sample subset for scrutiny.
  4. [§V-E / Fig. 6] The hyperparameters λ1 and λ2 are selected by varying them on IRCHN and inspecting R@1/AP. The paper defines only a training split and an evaluation split; there is no validation split. If the evaluation split was used for hyperparameter selection, the reported numbers are optimistically biased. Please define a validation protocol (e.g., a held-out portion of training locations, or cross-validation) and report the final test results using the selected hyperparameters.
  5. [§IV-D / Fig. 5] The inference-time descriptor is ambiguous. Fig. 5 labels the global branch 'for retrieval' and the relation branch 'for training,' but the ablations show that MAFE and BSSRM improve retrieval. If inference uses only the global descriptors, the contribution of MAFE/BSSRM is indirect (through regularizing the shared encoder); this should be stated explicitly. If inference uses the relation descriptors, the text in §IV-B and §IV-D that describes global descriptors as the retrieval feature must be corrected. Please specify the exact branch used at test time and how the relation branch influences it.
minor comments (5)
  1. [§V-B after Eq. (13)] The sentence 'The λ1 and λ1 in Eq (13) are both set to 1' contains a typo: the second λ1 should be λ2.
  2. [Table IV caption] The term 'individually trained protocol' is used but not defined in the text. Please define what 'individually trained' means here (e.g., each method is trained from scratch on IRCHN) and how it differs from the 'transfer' protocol mentioned in the introduction.
  3. [§III-E] The first sentence of §III-E starts with 'The evaluation set is divided...' which is fine, but the phrase 'visible and infrared drone images are separately used as queries' should clarify that this yields four settings overall; the sentence is slightly hard to parse.
  4. [Fig. 6] The figure caption says 'R@1 and AP for D→S and S→D retrieval under different values of (a)–(b) λ1 and (c)–(d) λ2,' but the plot axes and legends are not described in the text. Please ensure each panel is labeled clearly in the figure itself.
  5. [References] Some references are incomplete or have inconsistent formatting (e.g., [10] and [19] are arXiv preprints without a published venue; [1] and [26] are from 2025/2026). This is not a technical issue but should be cleaned for camera-ready.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: benchmark and method evaluation are self-contained; SOTRA is an auxiliary self-distillation loss, not a fitted prediction.

full rationale

The paper's empirical claims rest on a standard train/test protocol: IRCHN is a newly introduced benchmark with disjoint training and evaluation location splits, and MASTR-Net is trained on the training split and evaluated on held-out query/gallery locations. No parameter is fitted to the test-set labels and then reported as a prediction. The closest step to a circular-looking construction is the SOTRA loss, which trains the relation branch to match Sinkhorn-normalized soft targets derived from the global branch's similarity matrix. However, this is explicitly a training-time auxiliary objective; retrieval is performed using the global branch, which is independently supervised by an InfoNCE loss against geographic location labels. The soft target T is derived from the same network's global descriptors, but this is self-distillation rather than a circular derivation of the paper's central claim. The paper also cites several works by its own authors (e.g., MEAN [1], DMNIL [4], CDIKTNet [13]), but those citations are used as baselines or related-work context, not as load-bearing justification for MASTR-Net's design or for the benchmark's validity. There is no invoked uniqueness theorem, no fitted law renamed as a prediction, and no ansatz smuggled in via a self-citation. The skeptical concerns about uncontrolled baseline backbones, single-run results, and the absence of released benchmark data are threats to external validity or reproducibility, but they are not instances of circular derivation under the criteria used here. Consequently, the derivation chain is self-contained and no significant circularity is identified.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

MASTR-Net's central empirical claim rests on a set of hand-set/tuned hyperparameters and on the geographic-identity alignment of the new benchmark; no new physical entities are introduced.

free parameters (5)
  • λ1, λ2 loss weights = 1.0, 1.0
    Selected by hyperparameter sweep in Fig. 6 on the IRCHN evaluation set; used for all reported results.
  • τ_t (SOTRA target temperature) = 0.07
    Hand-set in Section V-B; controls sharpness of Sinkhorn target.
  • β (positive bias in SOTRA) = 2.0
    Hand-set to reinforce paired samples in Sinkhorn target.
  • K (Sinkhorn iterations) = 5
    Hand-set; number of alternating normalizations.
  • ρ_c (MAFE modulation range) = 0.10
    Hand-set in Section IV-B to bound channel-wise rescaling to (0.9, 1.1).
assumptions (4)
  • standard math Sinkhorn normalization yields a valid doubly-stochastic soft target.
    Used in Eq. (11); relies on Cuturi (2013).
  • domain assumption Geographic identity (GPS coordinate) is a reliable positive-pair label for retrieval.
    Section III-B/E: visible, infrared, and satellite images with the same location ID are treated as matched; no pixel-level alignment is performed.
  • domain assumption A single shared ConvNeXt encoder can represent visible, infrared, and satellite images in a common space.
    Section IV-A: the same F_b encodes all three modalities.
  • domain assumption Infrared and visible images captured at one location depict the same geographic content.
    Section III-B: pairs are collected together and associated by coordinates; temporal or seasonal changes are not discussed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Unified Benchmark and Modality-Adaptive Network for Day-and-Night Drone-View Geo-Localization." pith.science (2026). https://pith.science/paper/XBJRN4ZH

@misc{pith2026260725778,
  author       = {Pith},
  title        = {Pith review of: A Unified Benchmark and Modality-Adaptive Network for Day-and-Night Drone-View Geo-Localization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XBJRN4ZH}},
  note         = {Machine review of arXiv:2607.25778}
}
read the original abstract

Most existing drone-view geo-localization (DVGL) benchmarks contain drone imagery captured under a single illumination condition and lack geographically aligned visible drone images, infrared drone images, and satellite images from the same locations. To evaluate the generalization capability of DVGL methods under challenging illumination conditions, some methods train models on a visible benchmark and test them on an independent infrared benchmark. This protocol essentially constitutes transfer between datasets, which makes it difficult to systematically evaluate DVGL across daytime and nighttime conditions within a unified benchmark. To address this limitation, we construct IRCHN,a real-world DVGL benchmark designed for localization across different illumination conditions. IRCHN contains 26,460 images collected from 8,820 geographic locations across four representative scene categories, including farmland, coastline, forest, and urban areas. Each location provides one visible drone image, one infrared drone image, and one corresponding satellite image, which enables unified evaluation of DVGL methods across different illumination conditions and sensing modalities. We further propose the Modality-Adaptive State-Space Transport Relation Network (MASTR-Net), a DVGL framework tailored to localization under varying illumination conditions. MASTR-Net integrates modality-adaptive feature enhancement, bidirectional selective state-space relation modeling, and soft optimal transport relation alignment to jointly reduce modality gaps and view-induced structural discrepancies. Extensive experiments demonstrate that MASTR-Net outperforms existing state-of-the-art methods on IRCHN for localization under varying illumination conditions and achieves competitive performance on two infrared benchmarks, IR-VL328 and CVGL-RGBT. Code: https://github.com/SongtianhaoXu/MASTR-Net

Figures

Figures reproduced from arXiv: 2607.25778 by the authors.

Figure 1
Figure 1. Motivation for daytime and nighttime DVGL. (a) Localization from visible drone imagery is mainly affected by the large viewpoint difference from satellite references, while aligned infrared observations are unavailable. (b) Localization from infrared drone imagery is affected by both viewpoint and modality differences, while the corresponding visible observations are missing. The lack of geographically aligned visib… view at source ↗
Figure 2
Figure 2. Overview of the proposed IRCHN benchmark. IRCHN provides geo-location-aligned visible-light drone, infrared drone, and satellite images across cropland, urban, coastal, and forest scenes. Its operational acquisition supports unified day–night localization and evaluation of cross-view and cross￾modal discrepancies under diverse real-world conditions. [10]. However, this protocol essentially measures transfer between … view at source ↗
Figure 3
Figure 3. Representative samples from IRCHN. Each row presents one scene category, and the columns show geographically corresponding visible drone, infrared drone, and satellite images. The images share the same location identity but differ substantially in viewpoint, resolution, and appearance. TABLE III SCENE-WISE STATISTICS OF IRCHN. DRONE COUNTS INCLUDE BOTH VISIBLE AND INFRARED IMAGES. Scene Training Query Gallery Drone … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Overall pipeline of the proposed MASTR-Net. (a) A shared ConvNeXt-Base encoder extracts part features f d,f s and global descriptors g d, g s . (b) MAFE generates enhanced part features ef d,ef s through local enhancement, channel modulation, and residual fusion. (c) B…
Figure 5
Figure 5. Figure 5: Matching and supervision strategy of SOTRA. The normalized global and relation descriptors form the similarity matrices Rg and Rr, respectively. Sinkhorn balancing converts Rg into the soft target T, while Rr produces the bidirectional distributions P and Pe. Their ali…
Figure 6
Figure 6. Figure 6: Sensitivity to the loss balancing coefficients on IRCHN. R@1 and AP for D→S and S→D retrieval under different values of (a)–(b) λ1 and (c)–(d) λ2 on the visible and infrared modalities. D→S and S→D directions, respectively, demonstrating the benefit of structural relat…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 1 linked inside Pith

  1. [1]

    Multilevel embedding and alignment network with consistency and invariance learning for cross- view geo-localization,

    Z. Chen, Z.-X. Yang, and H.-J. Rong, “Multilevel embedding and alignment network with consistency and invariance learning for cross- view geo-localization,”IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1–15, 2025

    2025

  2. [2]

    Enhancing cross-view geo-localization with domain alignment and scene consistency,

    P. Xia, Y . Wan, Z. Zheng, Y . Zhang, and J. Deng, “Enhancing cross-view geo-localization with domain alignment and scene consistency,”IEEE Transactions on Circuits and Systems for Video Technology, pp. 1–12, 2024

    2024

  3. [3]

    MCCG: A convnext- based multiple-classifier method for cross-view geo-localization,

    T. Shen, Y . Wei, L. Kang, S. Wan, and Y .-H. Yang, “MCCG: A convnext- based multiple-classifier method for cross-view geo-localization,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 3, pp. 1456–1468, 2023

    2023

  4. [4]

    Without paired la- beled data: End-to-end self-supervised learning for drone-view geo- localization,

    Z. Chen, Z.-X. Yang, H.-J. Rong, and G. Li, “Without paired la- beled data: End-to-end self-supervised learning for drone-view geo- localization,”IEEE Transactions on Neural Networks and Learning Systems, pp. 1–15, 2026

    2026

  5. [5]

    Sample4Geo: Hard negative sampling for cross-view geo-localisation,

    F. Deuser, K. Habel, and N. Oswald, “Sample4Geo: Hard negative sampling for cross-view geo-localisation,” inIEEE/CVF International Conference on Computer Vision, 2023, pp. 16 847–16 856

    2023

  6. [6]

    Efficient spike-driven transformer for high-performance drone-view geo-localization,

    Z. Chen, H.-J. Rong, L. Xu, J. Lu, Z. Zou, Z.-X. Yang, and G. Li, “Efficient spike-driven transformer for high-performance drone-view geo-localization,”IEEE Transactions on Circuits and Systems for Video Technology, pp. 1–1, 2026

    2026

  7. [7]

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

    Z. Zheng, Y . Wei, and Y . Yang, “University-1652: A multi-view multi- source benchmark for drone-based geo-localization,” inACM Interna- tional Conference on Multimedia, no. 9, 2020, pp. 1395–1403

    2020

  8. [8]

    SUES-200: A multi-height multi-scene cross-view image benchmark across drone and satellite,

    R. Zhu, L. Yin, M. Yang, F. Wu, Y . Yang, and W. Hu, “SUES-200: A multi-height multi-scene cross-view image benchmark across drone and satellite,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 9, pp. 4825–4839, 2023

    2023

Show all 31 references
  1. [9]

    Vision- based UA V self-positioning in low-altitude urban environments,

    M. Dai, E. Zheng, Z. Feng, L. Qi, J. Zhuang, and W. Yang, “Vision- based UA V self-positioning in low-altitude urban environments,”IEEE Transactions on Image Processing, vol. 33, pp. 493–508, 2024

    2024

  2. [10]

    Object detection as an optional basis: A graph matching network for cross-view uav localization,

    T. Liu, K. Ren, and Q. Chen, “Object detection as an optional basis: A graph matching network for cross-view uav localization,”arXiv preprint arXiv:2511.02489, 2025

    2025

  3. [11]

    Cdm-net: A framework for cross- view geo-localization with multimodal data,

    X. Zhou, X. Yang, and Y . Zhang, “Cdm-net: A framework for cross- view geo-localization with multimodal data,”IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1–16, 2025

    2025

  4. [12]

    Geo2: Geometry-guided cross-view geo-localization and image synthesis,

    Y . Zhang, X. Zhang, G. Sun, Z. Lyu, S. Wshah, and C. Chen, “Geo2: Geometry-guided cross-view geo-localization and image synthesis,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 19 432–19 442

    2026

  5. [13]

    From limited labels to open domains: An efficient learning method for drone-view geo-localization,

    Z. Chen, Z.-X. Yang, H.-J. Rong, J. Lang, and G. Li, “From limited labels to open domains: An efficient learning method for drone-view geo-localization,”IEEE Transactions on Multimedia, pp. 1–13, 2026

    2026

  6. [14]

    Predicting ground- level scene layout from aerial imagery,

    M. Zhai, Z. Bessinger, S. Workman, and N. Jacobs, “Predicting ground- level scene layout from aerial imagery,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 867– 875. 11

    2017

  7. [15]

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

    L. Liu and H. Li, “Lending orientation to neural networks for cross- view geo-localization,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 5624–5633

    2019

  8. [16]

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

    S. Zhu, T. Yang, and C. Chen, “Vigor: Cross-view image geo-localization beyond one-to-one retrieval,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2021, pp. 3640–3649

    2021

  9. [17]

    Multi-object tracking meets moving uav,

    S. Liu, X. Li, H. Lu, and Y . He, “Multi-object tracking meets moving uav,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 8876–8885

    2022

  10. [18]

    Must: The first dataset and unified framework for multispectral uav single object track- ing,

    H. Qin, T. Xu, T. Li, Z. Chen, T. Feng, and J. Li, “Must: The first dataset and unified framework for multispectral uav single object track- ing,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 16 882–16 891

    2025

  11. [19]

    Uav- visloc: A large-scale dataset for uav visual localization,

    W. Xu, Y . Yao, J. Cao, Z. Wei, C. Liu, J. Wang, and M. Peng, “Uav- visloc: A large-scale dataset for uav visual localization,”arXiv preprint arXiv:2405.11936, 2024

    2024 arXiv

  12. [20]

    Game4loc: A uav geo-localization benchmark from game data,

    Y . Ji, B. He, Z. Tan, and L. Wu, “Game4loc: A uav geo-localization benchmark from game data,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 4, 2025, pp. 3913–3921

    2025

  13. [21]

    Uav- geoloc: A large-vocabulary dataset and geometry-transformed method for uav geo-localization,

    R. Wu, J. Deng, M. Mou, X. He, M. Zhang, Y . Liu, and S. Yan, “Uav- geoloc: A large-vocabulary dataset and geometry-transformed method for uav geo-localization,”IEEE Robotics and Automation Letters, vol. 10, no. 9, pp. 8978–8985, 2025

    2025

  14. [22]

    Each part matters: Local patterns facilitate cross-view geo-localization,

    T. Wang, Z. Zheng, C. Yan, J. Zhang, Y . Sun, B. Zheng, and Y . Yang, “Each part matters: Local patterns facilitate cross-view geo-localization,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 2, pp. 867–879, 2021

    2021

  15. [23]

    A transformer-based fea- ture segmentation and region alignment method for UA V-view geo- localization,

    M. Dai, J. Hu, J. Zhuang, and E. Zheng, “A transformer-based fea- ture segmentation and region alignment method for UA V-view geo- localization,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 7, pp. 4376–4389, 2021

    2021

  16. [24]

    CAMP: Across-view geo-localization method using contrastive attributes mining and position-aware partitioning,

    Q. Wu, Y . Wan, Z. Zheng, Y . Zhang, G. Wang, and Z. Zhao, “CAMP: Across-view geo-localization method using contrastive attributes mining and position-aware partitioning,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–14, 2024

    2024

  17. [25]

    Mfrgn: Multi-scale feature representation generalization network for ground-to-aerial geo- localization,

    Y . Wang, J. Zhang, R. Wei, W. Gao, and Y . Wang, “Mfrgn: Multi-scale feature representation generalization network for ground-to-aerial geo- localization,” inProceedings of the 32nd ACM International Conference on Multimedia(ACMMM), 2024, pp. 2574–2583

    2024

  18. [26]

    Surfnet: A surface-aware uav–satellite geoloca- tion framework via feature aggregation and dual positional encoding,

    K. Liu and W. Zhang, “Surfnet: A surface-aware uav–satellite geoloca- tion framework via feature aggregation and dual positional encoding,” IEEE Transactions on Geoscience and Remote Sensing, vol. 64, pp. 1– 15, 2026

    2026

  19. [27]

    Unsu- pervised multiview uav image geolocalization via iterative rendering,

    H. Li, C. Xu, W. Yang, L. Mi, H. Yu, H. Zhang, and G.-S. Xia, “Unsu- pervised multiview uav image geolocalization via iterative rendering,” IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1– 15, 2025

    2025

  20. [28]

    Uniabg: Unified adversarial view bridging and graph correspondence for unsupervised cross-view geo-localization,

    C. Chen, Q. Chen, B. Yang, and X. Zhang, “Uniabg: Unified adversarial view bridging and graph correspondence for unsupervised cross-view geo-localization,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 4, 2026, pp. 2823–2831

    2026

  21. [29]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 11 976–11 986

    2022

  22. [30]

    Eca-net: Efficient channel attention for deep convolutional neural networks,

    Q. Wang, B. Wu, P. Zhu, P. Li, W. Zuo, and Q. Hu, “Eca-net: Efficient channel attention for deep convolutional neural networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11 534–11 542

    2020

  23. [31]

    Sinkhorn distances: Lightspeed computation of optimal transport,

    M. Cuturi, “Sinkhorn distances: Lightspeed computation of optimal transport,” inAdvances in Neural Information Processing Systems, vol. 26, 2013, pp. 2292–2300

    2013

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.