Pith. sign in

REVIEW 4 major objections 5 minor 51 references

Dynamic Contrastive Learning for Hierarchical Retrieval: A Case Study of Distance-Aware Cross-View Geo-Localization

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims cross-view geo-localization should rank candidates by geographic distance, and that a contrastive loss with decreasing margins—DyCL—makes hierarchical retrieval work across spatial scales.

desk verdict A genuinely new benchmark and task formulation, but the paper's advertised margin mechanism doesn't logically enforce the claimed distance-aware ranking. read the letter →

arxiv 2506.23077 v1 pith:JOKZYZML submitted 2025-06-29 cs.CV

Suofei Zhang , Xinxin Wang , Xiaofu Wu , Quan Zhou , Haifeng Hu This is my paper · ORCID
classification cs.CV
keywords cross-viewgeo-localizationhierarchicalretrievalcontrastivelearningdistance-awarerankingdynamicmetricbenchmarkre-ranking
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

Cross-view geo-localization is usually treated as exact-match retrieval: the goal is to put the image of the same building at the top of the ranking. This paper argues that the ranking should instead be distance-aware, so that when the exact building is missed, the top results are buildings that are geographically close to the query and may carry contextual clues about it. To make this precise, the authors introduce Distance-Aware Cross-View Geo-Localization (DACVGL), build the DA-Campus benchmark with GPS distance annotations for 750 university buildings, and show that conventional metric learning fails on it because adjacent spatial scales pull the embedding in opposite directions. Their remedy, Dynamic Contrastive Learning (DyCL), enforces decreasing similarity margins across scales and, combined with HAPPIER and a multi-scale re-ranking step, improves hierarchical average precision on Satellite to Drone retrieval from 45.00 to 49.97 over the strongest single-scale baseline.

What carries the argument

The load-bearing mechanism is the DyCL loss in Eq. (4): for each anchor, the similarity to images inside a scale $S_c^{\le l}$ is pushed above the similarity to pure negatives (buildings farther than the largest threshold) by a margin $m_l$, with margins strictly decreasing as scales grow. An anchor-specific, nested definition of scales $S_c = \{S_c^l\}$ replaces any fixed semantic hierarchy, because in DACVGL spatial relevance is defined continuously around each building. A second component, Multi-Scale Re-ranking, applies k-reciprocal re-ranking with scale-dependent neighborhood sizes $k_l$ and accumulates results with a mask so that finer-scale orderings are not disturbed.

What would settle it

Re-geocode all 750 buildings with survey-grade positioning and recompute the three relevance bins; if even a few percent of building pairs change bins, the hierarchical labels shift, and the reported H-AP gains of DyCL could be an artifact of label noise. A cleaner test is to evaluate on real drone imagery with known positions and check whether the top-ranked failures are actually closer to the query than those of single-scale baselines.

Watch

Extended reading notes

Core claim

Cross-view geo-localization is usually judged by exact-match retrieval: does the top-ranked reference image depict the same building as the query. The paper argues this misses the geographic context that matters when exact matches fail, and replaces it with Distance-Aware Cross-View Geo-Localization (DACVGL), a hierarchical retrieval problem in which relevance is defined by physical distance from the query building. The central discovery is that this problem cannot be solved by ordinary metric learning because adjacent spatial scales impose conflicting objectives on the same embedding space. Their Dynamic Contrastive Learning (DyCL) loss resolves the conflict by enforcing margins $m_0 > m_1 > \cdots > 0$ between positive and negative similarities at each scale, with distances partitioned at 0, 200, and 500 m. On the new DA-Campus benchmark, DyCL combined with HAPPIER and a multi-scale re-ranking step lifts hierarchical average precision on Satellite to Drone from 45.00 (the strongest single-scale baseline, CV-Cities) to 49.97.

Load-bearing premise

The benchmark and the loss both assume that GPS-derived Euclidean distances between building coordinates, binned at 0, 200, and 500 meters, are the correct ground truth for spatial relevance; if the geocodes are imprecise or the fixed thresholds don't match a campus's layout, every distance label, margin, and reported metric is unreliable.

Editorial extensions

If this is right

  • Trained models produce rankings in which, even when the exact building is missed, the top results are geographically closer to the query and more likely to contain contextual cues about it.
  • Combining DyCL with HAPPIER consistently beats both the best single-scale baseline (CV-Cities) and each method alone across hierarchical metrics on both retrieval directions.
  • Multi-scale re-ranking generalizes beyond geo-localization: it improves mAP on standard hierarchical retrieval datasets (DyML-Vehicle, DyML-Animal, DyML-Product) without hurting R@1.
  • The DA-Campus benchmark supplies a reusable evaluation protocol (R@K, mAP, H-AP, ASI, NDCG) with distance-based relevance levels for future methods.

Reading between the lines

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

  • The same decreasing-margin idea could transfer to any retrieval task with continuous relevance, such as place recognition with graded pose error, where fixed semantic bins are equally arbitrary.
  • Because DA-Campus is built from synthetic drone renders and Wikipedia geocodes, a natural next test is real drone imagery with surveyed positions; the fixed 0/200/500 m thresholds may need to become per-location or learned.
  • The binned formulation discards ordering within a bin; a continuous margin proportional to distance is a natural extension that might further improve fine-grained spatial ranking.
  • The results suggest the remaining weakness is small-scale R@1, where single-scale classification still leads; a hybrid that preserves fine-grained discrimination while adding hierarchical margins could close that gap.
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

4 major / 5 minor

Summary. The paper introduces a distance-aware cross-view geo-localization (DACVGL) task, constructs the DA-Campus benchmark with multi-view (satellite/drone) imagery and GPS-derived distance annotations, and proposes Dynamic Contrastive Learning (DyCL), a multi-scale contrastive loss with decreasing margins across three distance levels. It also proposes Multi-Scale Re-ranking (MSRerank). Experiments on DA-Campus report that DyCL, especially combined with HAPPIER and MSRerank, improves hierarchical retrieval metrics and overall geo-localization accuracy over several baselines. The code and dataset are promised to be public.

Significance. If the claims hold, the paper would provide a useful new benchmark for distance-aware cross-view retrieval and a training loss that improves ranking quality at multiple spatial scales. Strengths include the construction and release of DA-Campus, the breadth of baseline comparisons in Tables II and III, and the demonstration that MSRerank transfers to standard hierarchical retrieval datasets (Table IV). However, the central claim that DyCL enforces a monotone distance-aware hierarchy is not supported by the loss as written: the decreasing margins are lower bounds against a single far-negative set, not constraints between adjacent distance rings. The main mechanism is also incompletely specified because the margin values m_l are never given. These issues are load-bearing for the paper's core contribution.

major comments (4)
  1. [Section IV-C, Eq. (3)-(4)] The loss in Eq. (4) does not enforce the claimed monotone hierarchy. For each scale l, positives are drawn from the cumulative set S_c^{<=l} and negatives only from the "pure negative" set beyond the largest threshold. Thus the loss imposes r(S_c^{<=l}) - r(far) >= m_l for l=0,1,2, but it never compares a same-building sample against a 0-200 m sample, nor a 0-200 m sample against a 200-500 m sample. Decreasing margins m_0 > m_1 > m_2 do not imply r(same) > r(0-200 m) > r(200-500 m) > r(far): for instance, r(same)-r(far)=0.6 and r(200-500 m)-r(far)=0.1 can both hold while the 200-500 m building is ranked above the exact match. Consequently the statement that DyCL "progressively aligns feature representations according to hierarchical spatial margins" is not a logical consequence of the loss, and the H-AP gains in Table II could be driven by separating all in-threshold buildings from far negatives rather than by true distance-ordered ranking. Please modify the loss to include inter-ring negatives, or provide direct evidence that learned similarities are monotone in geographic distance, and adjust the claims accordingly.
  2. [Section IV-A vs. Section IV-C] There is an inconsistency between the task formulation and the proposed loss. Eq. (2) defines the objective as maximizing r_ij - r_ik for all positive sets S_c^{<=l} against their complements S_c^{>l}, which includes adjacent rings as negatives. However, Eq. (4) restricts negatives to "pure negatives" (buildings beyond the largest threshold) only. The loss therefore optimizes a strictly weaker objective than Eq. (2), and the paper's statement that DyCL addresses the competing objectives in Eq. (2) is not justified. Please clarify whether Eq. (2) is intended as an idealized goal and explain why restricting negatives to the far set is sufficient, or align the loss with Eq. (2).
  3. [Section IV-C, hyperparameters] The margin values m_l, which are the core mechanism of DyCL, are never specified. The text only says "we set m_0 > m_1 > ... > m_{L-1} > 0" but gives no actual values, no schedule, and no sensitivity analysis. Without these values, the reported DyCL results cannot be reproduced and the claim that decreasing margins are what drives the improvement is untestable. Please provide the exact margin schedule used in Tables II and III, and a small ablation over margin schedules.
  4. [Section III-A, III-C and V-E] The validity of the DA-Campus ground truth is not established. The paper states that building names from Wikipedia were "geocoded into precise geographic coordinates" but reports no geocoding error analysis, and the distance thresholds (0, 200, 500 m) are fixed without examining whether they are appropriate for the density of campus buildings. Since every distance label, every training margin, and every evaluation metric depends on these coordinates and thresholds, the benchmark's reliability rests on an unquantified assumption. Please report geocoding accuracy, building-density statistics, and a sensitivity analysis over threshold choices.
minor comments (5)
  1. [Section V-E] Hyperparameters tau and lambda_1 are selected by ablations on DA-Campus itself, and no error bars or multiple-seed variances are reported anywhere in Tables II or III. At minimum, please report the variance over at least three training seeds for the main comparisons.
  2. [Section IV-D, Algorithm 1] The title of Algorithm 1 reads "Multl-Scale Re-ranking Algorithm"; the typo should be corrected to "Multi-Scale".
  3. [Section IV-D] The sentence "It and can be seamlessly plugged into various cross-domain and standard hierarchical retrieval frameworks" is grammatically incomplete and should be rewritten.
  4. [Section V-E] "Base on the results" should be "Based on the results", and in the conclusion "More validate" should be "More validation" or similar.
  5. [Section IV-C, Eq. (4)] The set S_c^L used for pure negatives is not defined in the text; the paper earlier defines S_c^{>l} as the complement of S_c^{<=l}. Please define S_c^L explicitly or replace it with S_c^{>L-1} throughout.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: DyCL is not a re-derivation of H-AP and no fitted value is renamed as a prediction; the unsupported hierarchy-enforcement claim in Eq. (3)-(4) is a correctness gap, not a circular one.

full rationale

No load-bearing circularity was found. The proposed DyCL loss (Section IV-C, Eq. (4)) is not a re-derivation of the evaluation metric: it is a margin-based contrastive loss separating cumulative positive sets S^{<=l}_c from the beyond-largest-threshold set at each scale, whereas H-AP, ASI, and NDCG are external metrics (H-AP from [22]) computed at test time, and the marginal DyCL gain is measured against an HAPPIER baseline that also optimizes H-AP. The margins m_l are stated intuitively (m_0 > m_1 > ... > 0) and are not fitted to the test set, so there is no fitted-input-called-prediction step; the MSRerank neighborhood sizes k_l are selected from training statistics via Eq. (6) with mu = 0.1 and are additionally validated on the external DyML-Vehicle, DyML-Animal, and DyML-Product benchmarks in Table IV, which provides independent grounding. Self-citations are not load-bearing: the dynamic metric learning framework [21], [22] is external prior work, and no uniqueness theorem or ansatz is imported from the present authors' earlier papers. Three concerns raised by the skeptical reading are real but belong to correctness and validity, not circularity. First, Eq. (3)-(4) only constrain each cumulative positive set against the >500 m negative set and never compare adjacent rings (e.g., 0-200 m vs 200-500 m), so decreasing margins do not logically imply r(same) > r(0-200m) > r(200-500m) > r(far); the claim that DyCL 'progressively aligns feature representations according to hierarchical spatial margins' (Section IV-C) is therefore an over-claim relative to the loss, and H-AP gains could in principle come from same-vs-far separation alone. That is a broken derivation, not an identity of input and output. Second, the headline 49.97 H-AP (Table II) is achieved by DyCL+HAPPIER+MSRerank, a combination containing HAPPIER, an external H-AP-optimizing method; however, the paper reports the HAPPIER-only baseline (44.95) and the DyCL-only result (46.02, above the strongest single-scale CV-Cities at 45.00), so the proposed contribution is separable and the comparison is not forced by construction. Third, DA-Campus is a self-constructed benchmark whose distance labels come from Wikipedia geocodes asserted as 'precise GPS tags' (Section III-A) without error analysis; any systematic geocoding bias at the tens-of-meters scale would corrupt the 200 m and 500 m thresholds shared by the loss, the relevance labels, and every reported metric.

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

The central empirical claims rest on a small set of hand-chosen thresholds, margins, and loss weights. None are derived, and the margin values are absent from the text. No new physical entities are introduced.

free parameters (5)
  • Distance thresholds defining relevance levels = 0 m, 200 m, 500 m
    Chosen thresholds in Section III-C; they define the hierarchy used both in training and evaluation, with no sensitivity analysis reported.
  • Margin schedule m_l in Eq. (3) = not reported
    The DyCL loss depends on decreasing margins m_0 > m_1 > m_2 > 0, but their numerical values are never given, so the core training mechanism is underspecified.
  • Contrastive temperature tau = 32
    Selected by ablation in Section V-E on DA-Campus; no validation split is described.
  • Loss weights lambda_1, lambda_2, lambda_3 = 0.2, 0.1, 0.9
    Selected based on ablation on the Satellite to Drone task in Section V-E; the choice is then evaluated on the same benchmark reported in Table II.
  • Re-ranking scale factor mu = 0.1
    Empirical hyperparameter in Eq. (6) for MSRerank, fixed without sensitivity analysis.
assumptions (4)
  • domain assumption GPS-derived Euclidean distance between building coordinates is a valid measure of spatial relevance for retrieval.
    DA-Campus ground truth and all evaluation metrics are built from pairwise GPS distances (Section III-B); geocoding errors would invalidate relevance labels.
  • domain assumption Synthetic drone views from Google Earth 3D models and satellite views from Google Maps provide enough cross-view correspondence to learn transferable features.
    Dataset construction in Section III-A replaces real drone flights with simulated views, an unverified proxy for real drone imagery.
  • ad hoc to paper Fixed distance thresholds (0, 200, 500 m) create a meaningful hierarchy for all buildings regardless of campus density.
    The hierarchy in Eq. (1) and all metrics in Section III-C depend on these thresholds; no adaptive or sensitivity analysis is provided.
  • ad hoc to paper A monotone decreasing margin schedule m_0 > m_1 > ... > 0 resolves the conflicting objectives between adjacent scales.
    The claim in Section IV-C that decreasing margins mitigate the contradiction in Eq. (2) is asserted intuitively and never proven or numerically specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic Contrastive Learning for Hierarchical Retrieval: A Case Study of Distance-Aware Cross-View Geo-Localization." pith.science (2026). https://pith.science/paper/JOKZYZML

@misc{pith2026250623077,
  author       = {Pith},
  title        = {Pith review of: Dynamic Contrastive Learning for Hierarchical Retrieval: A Case Study of Distance-Aware Cross-View Geo-Localization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JOKZYZML}},
  note         = {Machine review of arXiv:2506.23077}
}
read the original abstract

Existing deep learning-based cross-view geo-localization methods primarily focus on improving the accuracy of cross-domain image matching, rather than enabling models to comprehensively capture contextual information around the target and minimize the cost of localization errors. To support systematic research into this Distance-Aware Cross-View Geo-Localization (DACVGL) problem, we construct Distance-Aware Campus (DA-Campus), the first benchmark that pairs multi-view imagery with precise distance annotations across three spatial resolutions. Based on DA-Campus, we formulate DACVGL as a hierarchical retrieval problem across different domains. Our study further reveals that, due to the inherent complexity of spatial relationships among buildings, this problem can only be addressed via a contrastive learning paradigm, rather than conventional metric learning. To tackle this challenge, we propose Dynamic Contrastive Learning (DyCL), a novel framework that progressively aligns feature representations according to hierarchical spatial margins. Extensive experiments demonstrate that DyCL is highly complementary to existing multi-scale metric learning methods and yields substantial improvements in both hierarchical retrieval performance and overall cross-view geo-localization accuracy. Our code and benchmark are publicly available at https://github.com/anocodetest1/DyCL.

Figures

Figures reproduced from arXiv: 2506.23077 by the authors.

Figure 1
Figure 1. Illustration of a standard cross-view geo-localization retrieval process. Given a query image of a specific building marked in green, the system ranks [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Spatial annotation schemes. (a) Conventional annotation scheme in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the detailed spatial partitioning strategy in DACVGL. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Overview of the baseline cross-view geo-localization framework. Images from drone and satellite views are processed by a shared-weight backbone, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Illustration of the DyCL margin control mechanism across multiple geographic scales. Given an anchor sample [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Comparison of performance among single-scale models and a multi [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Distribution of position shifts across different ranking positions after [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Analysis of hyper-parameters τ and λ. For simplicity, only the results for the Satellite → Drone task are shown. Similar trends are observed for the symmetric Drone → Satellite task. (a) Impact of scaling factor τ in Eq. (4). (b) Impact of loss weight λ1 in Eq. (7). λ1…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 38 canonical work pages

  1. [1]

    Learning deep represen- tations for ground-to-aerial geolocalization,

    T.-Y . Lin, Y . Cui, S. Belongie, and J. Hays, “Learning deep represen- tations for ground-to-aerial geolocalization,” in2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 5007– 5015

    work page 2015

  2. [2]

    Cross-view image matching for geo-localization in urban environments,

    Y . Tian, C. Chen, and M. Shah, “Cross-view image matching for geo-localization in urban environments,” in2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1998– 2006

    work page 2017

  3. [3]

    24/7 place recognition by view synthesis,

    A. Torii, R. Arandjelovi ´c, J. Sivic, M. Okutomi, and T. Pajdla, “24/7 place recognition by view synthesis,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 40, no. 2, pp. 257–271, 2018

    work page 2018

  4. [4]

    Building information modeling and classification by visual learning at a city scale,

    Q. Yu, C. Wang, B. Cetiner, S. Yu, F. Mckenna, E. Taciroglu, and K. Law, “Building information modeling and classification by visual learning at a city scale,” 2019

    work page 2019

  5. [5]

    Visdrone-det2021: The vision meets drone object detection challenge results,

    Y . Cao, Z. He, L. Wang, W. Wang, Y . Yuan, D. Zhang, J. Zhang, P. Zhu, L. Van Gool, J. Han, S. Hoi, Q. Hu, M. Liu, C. Cheng, F. Liu, G. Cao, G. Li, H. Wang, J. He, J. Wan, Q. Wan, Q. Zhao, S. Lyu, W. Zhao, X. Lu, X. Zhu, Y . Liu, Y . Lv, Y . Ma, Y . Yang, Z. Wang, Z. Xu, Z. Luo, Z. Zhang, Z. Zhang, Z. Li, and Z. Zhang, “Visdrone-det2021: The vision meets...

    work page 2021

  6. [6]

    Netvlad: Cnn architecture for weakly supervised place recognition,

    R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and J. Sivic, “Netvlad: Cnn architecture for weakly supervised place recognition,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 5297–5307

    2016

  7. [7]

    Object retrieval with large vocabularies and fast spatial matching,

    J. Philbin, O. Chum, M. Isard, J. Sivic, and A. Zisserman, “Object retrieval with large vocabularies and fast spatial matching,” in2007 IEEE Conference on Computer Vision and Pattern Recognition, 2007, pp. 1–8

    work page 2007

  8. [8]

    Progressive learning for person re-identification with one example,

    Y . Wu, Y . Lin, X. Dong, Y . Yan, W. Bian, and Y . Yang, “Progressive learning for person re-identification with one example,”IEEE Transac- tions on Image Processing, vol. 28, no. 6, pp. 2872–2881, 2019

    work page 2019

Show all 51 references
  1. [9]

    Ranking with local regression and global alignment for cross media retrieval,

    Y . Yang, D. Xu, F. Nie, J. Luo, and Y . Zhuang, “Ranking with local regression and global alignment for cross media retrieval,” inProceedings of the 17th ACM International Conference on Multimedia, ser. MM ’09. New York, NY , USA: Association for Computing Machinery, 2009, p....

    2009

  2. [10]

    Cosface: Large margin cosine loss for deep face recognition,

    H. Wang, Y . Wang, Z. Zhou, X. Ji, D. Gong, J. Zhou, Z. Li, and W. Liu, “Cosface: Large margin cosine loss for deep face recognition,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 5265–5274

    2018

  3. [11]

    Arcface: Additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, J. Yang, N. Xue, I. Kotsia, and S. Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,”IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol. 44, no. 10, pp. 5962–5979, 2022

    2022

  4. [12]

    Arcgeo: Localizing limited field-of-view images using cross-view matching,

    M. Shugaev, I. Semenov, K. Ashley, M. Klaczynski, N. Cuntoor, M. W. Lee, and N. Jacobs, “Arcgeo: Localizing limited field-of-view images using cross-view matching,” in2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024, pp. 208–217

    2024

  5. [13]

    Wide-area image geolocal- ization with aerial reference imagery,

    S. Workman, R. Souvenir, and N. Jacobs, “Wide-area image geolocal- ization with aerial reference imagery,” inIEEE International Conference on Computer Vision (ICCV), 2015, pp. 1–9, acceptance rate: 30.3%

    2015

  6. [14]

    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,” in2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 5617–5626. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11

    2019

  7. [15]

    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,”ACM Multimedia, 2020

    2020

  8. [16]

    Uavm’23: 2023 workshop on uavs in multimedia: Capturing the world from a new perspective,

    Z. Zheng, Y . Shi, T. Wang, J. Liu, J. Fang, Y . Wei, and T.-s. Chua, “Uavm’23: 2023 workshop on uavs in multimedia: Capturing the world from a new perspective,” inProceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 9715–9717

    2023

  9. [17]

    Cv-cities: Advancing cross- view geo-localization in global cities,

    G. Huang, Y . Zhou, L. Zhao, and W. Gan, “Cv-cities: Advancing cross- view geo-localization in global cities,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 18, pp. 1592– 1606, 2025

    2025

  10. [18]

    Sample4geo: Hard negative sam- pling for cross-view geo-localisation,

    F. Deuser, K. Habel, and N. Oswald, “Sample4geo: Hard negative sam- pling for cross-view geo-localisation,” in2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 16 801–16 810

    2023

  11. [19]

    Simple, effective and general: A new backbone for cross-view image geo-localization,

    Y . Zhu, H. Yang, Y . Lu, and Q. Huang, “Simple, effective and general: A new backbone for cross-view image geo-localization,” 2023. [Online]. Available: https://arxiv.org/abs/2302.01572

    2023 arXiv

  12. [20]

    Beyond geo-localization: Fine-grained orientation of street-view images by cross-view matching with satellite imagery,

    W. Hu, Y . Zhang, Y . Liang, Y . Yin, A. Georgescu, A. Tran, H. Kruppa, S.-K. Ng, and R. Zimmermann, “Beyond geo-localization: Fine-grained orientation of street-view images by cross-view matching with satellite imagery,” inProceedings of the 30th ACM International Conference ...

    2022

  13. [21]

    Dy- namic metric learning: Towards a scalable metric space to accommodate multiple semantic scales,

    Y . Sun, Y . Zhu, Y . Zhang, P. Zheng, X. Qiu, C. Zhang, and Y . Wei, “Dy- namic metric learning: Towards a scalable metric space to accommodate multiple semantic scales,” in2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 5389–5398

    2021

  14. [22]

    Hi- erarchical average precision training for pertinent image retrieval,

    E. Ramzi, N. Audebert, N. Thome, C. Rambour, and X. Bitot, “Hi- erarchical average precision training for pertinent image retrieval,” in Computer Vision – ECCV 2022, S. Avidan, G. Brostow, M. Ciss ´e, G. M. Farinella, and T. Hassner, Eds. Cham: Springer Nature Switzerland, 202...

    2022

  15. [23]

    Netvlad: Cnn architecture for weakly supervised place recognition,

    R. Arandjelovi ´c, P. Gronat, A. Torii, T. Pajdla, and J. Sivic, “Netvlad: Cnn architecture for weakly supervised place recognition,”IEEE Trans- actions on Pattern Analysis and Machine Intelligence, vol. 40, no. 6, pp. 1437–1451, 2018

    2018

  16. [24]

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

    S. Hu, M. Feng, R. M. H. Nguyen, and G. H. Lee, “Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2018, pp. 7258–7267

    2018

  17. [25]

    Optimal feature transport for cross-view image geo-localization,

    Y . Shi, X. Yu, L. Liu, T. Zhang, and H. Li, “Optimal feature transport for cross-view image geo-localization,” 2019. [Online]. Available: https://arxiv.org/abs/1907.05021

    2019 arXiv

  18. [26]

    Spatial-aware feature aggregation for image based cross-view geo-localization,

    Y . Shi, L. Liu, X. Yu, and H. Li, “Spatial-aware feature aggregation for image based cross-view geo-localization,” in Advances in Neural Information Processing Systems, H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch ´e-Buc, E. Fox, and R. Garnett, Eds., vol. 32. Curran ...

    2019

  19. [27]

    Where am i looking at? joint location and orientation estimation by cross-view matching,

    Y . Shi, X. Yu, D. Campbell, and H. Li, “Where am i looking at? joint location and orientation estimation by cross-view matching,” in2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 4063–4071

    2020

  20. [28]

    Localizing and orienting street views using overhead imagery,

    N. N. V o and J. Hays, “Localizing and orienting street views using overhead imagery,” inComputer Vision – ECCV 2016, B. Leibe, J. Matas, N. Sebe, and M. Welling, Eds. Cham: Springer International Publishing, 2016, pp. 494–509

    2016

  21. [29]

    Facenet: A unified embed- ding for face recognition and clustering,

    F. Schroff, D. Kalenichenko, and J. Philbin, “Facenet: A unified embed- ding for face recognition and clustering,” in2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 815–823

    2015

  22. [30]

    Hierarchical temporal modeling with mutual distance matching for video based person re- identification,

    P. Li, P. Pan, P. Liu, M. Xu, and Y . Yang, “Hierarchical temporal modeling with mutual distance matching for video based person re- identification,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 31, no. 2, pp. 503–511, 2021

    2021

  23. [31]

    Dimensionality reduction by learning an invariant mapping,

    R. Hadsell, S. Chopra, and Y . LeCun, “Dimensionality reduction by learning an invariant mapping,” in2006 IEEE Computer Society Con- ference on Computer Vision and Pattern Recognition (CVPR’06), vol. 2, 2006, pp. 1735–1742

    2006

  24. [32]

    Improved deep metric learning with multi-class n-pair loss objective,

    K. Sohn, “Improved deep metric learning with multi-class n-pair loss objective,” inAdvances in Neural Information Processing Systems, D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, Eds., vol. 29. Curran Associates, Inc., 2016. [Online]. Available: https://proceedin...

    2016

  25. [33]

    Unlabeled samples generated by gan improve the person re-identification baseline in vitro,

    Z. Zheng, L. Zheng, and Y . Yang, “Unlabeled samples generated by gan improve the person re-identification baseline in vitro,” in2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 3774– 3782

    2017

  26. [34]

    Dual-path convolutional image-text embeddings with instance loss,

    Z. Zheng, L. Zheng, M. Garrett, Y . Yang, M. Xu, and Y .-D. Shen, “Dual-path convolutional image-text embeddings with instance loss,” ACM Trans. Multimedia Comput. Commun. Appl., vol. 16, no. 2, May

  27. [35]

    On adaptive attacks to adversarial example defenses,

    F. Tramer, N. Carlini, W. Brendel, and A. Madry, “On adaptive attacks to adversarial example defenses,” inProceedings of the 34th International Conference on Neural Information Processing Systems, ser. NIPS ’20. Red Hook, NY , USA: Curran Associates Inc., 2020

    2020

  28. [36]

    Deepfashion: Powering robust clothes recognition and retrieval with rich annotations,

    Z. Liu, P. Luo, S. Qiu, X. Wang, and X. Tang, “Deepfashion: Powering robust clothes recognition and retrieval with rich annotations,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 1096–1104

    2016

  29. [37]

    Cumulated gain-based evaluation of ir techniques,

    K. J ¨arvelin and J. Kek ¨al¨ainen, “Cumulated gain-based evaluation of ir techniques,”ACM Trans. Inf. Syst., vol. 20, no. 4, p. 422–446, Oct

  30. [38]

    Classification is a strong baseline for deep metric learning,

    A. Zhai and H.-Y . Wu, “Classification is a strong baseline for deep metric learning,” 2019. [Online]. Available: https://arxiv.org/abs/1811.12649

    2019 arXiv

  31. [39]

    Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis,

    E. W. Teh, T. DeVries, and G. W. Taylor, “Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis,” inComputer Vision – ECCV 2020, A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, Eds. Cham: Springer International Publishing, 2020, pp. 448–464

    2020

  32. [40]

    Re-ranking person re- identification with k-reciprocal encoding,

    Z. Zhong, L. Zheng, D. Cao, and S. Li, “Re-ranking person re- identification with k-reciprocal encoding,” in2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 3652– 3661

    2017

  33. [41]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    2016

  34. [42]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255

    2009

  35. [43]

    Representation learning with contrastive predictive coding,

    A. van den Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” 2019. [Online]. Available: https: //arxiv.org/abs/1807.03748

    2019 arXiv

  36. [44]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” inProceedings of the 38th International Conference on Machine Le...

    2021

  37. [45]

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

    X. Wang, X. Han, W. Huang, D. Dong, and M. R. Scott, “Multi- similarity loss with general pair weighting for deep metric learning,” in2019 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), 2019, pp. 5017–5025

    2019

  38. [46]

    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, 2022

    2022

  39. [47]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut, A. ...

    2024 arXiv

  40. [48]

    A new measure of rank correlation,

    M. G. KENDALL, “A new measure of rank correlation,”Biometrika, vol. 30, no. 1-2, pp. 81–93, 06 1938. [Online]. Available: https: //doi.org/10.1093/biomet/30.1-2.81

    1938 doi

  41. [49]

    Sampling matters in deep embedding learning,

    C.-Y . Wu, R. Manmatha, A. J. Smola, and P. Kr ¨ahenb¨uhl, “Sampling matters in deep embedding learning,” in2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2859–2867

    2017

  42. [2002]

    Available: https://doi.org/10.1145/582415.582418

    [Online]. Available: https://doi.org/10.1145/582415.582418

  43. [2020]

    Available: https://doi.org/10.1145/3383184

    [Online]. Available: https://doi.org/10.1145/3383184

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.