Pith. sign in

REVIEW 2 major objections 3 minor 46 references

RIM: A Retrieval-In-Matching Framework for Cross-Domain Global Visual Localization of UAVs

T0 review · 2 major / 3 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 The paper claims that a single frozen DINOv2-B token field can serve both global retrieval and local matching for UAV localization, cutting per-query latency to 67.9 ms while keeping re-ranking Recall@1 within about 1 percentage point of a

desk verdict Solid, transparent systems paper with a genuinely shared-representation architecture; just define the 'Urbanscape-OOP' split before I believe the zero-shot headline. read the letter →

arxiv 2607.20116 v1 pith:I5PASVAP submitted 2026-07-22 cs.CV

Xin Li , Siyuan Duan , Shang Wang , Zhimin Mao , Bingliang Hu , Geng Zhang This is my paper
classification cs.CV
keywords UAVglobalvisuallocalizationplacerecognitionretrieval-in-matchingDINOv2SALADlocalfeaturematchingcross-domainadaptation6-DoFposeestimation
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 claims that a UAV can be localized against a rendered 3D-mesh reference map using a single shared representation for both retrieval and matching. It proposes RIM, which freezes a DINOv2-B backbone fine-tuned for cross-domain retrieval and distills a local-descriptor decoder on top of the same token field. Because the retrieval head is frozen, the global descriptor is unchanged by construction; the decoder only learns to interpret the shared features for geometric verification. On two zero-shot benchmarks, the two-stage fine-tuning raises Recall@1 over vanilla SALAD by up to 13.77 percentage points, and the complete query—retrieval, matching, and pose estimation—runs in 67.9 ms. If correct, the result makes 6-DoF visual localization practical for GNSS-denied UAVs on edge hardware.

What carries the argument

The load-bearing identity is g_after(I) = g_before(I) (Eq. 8): the frozen SALAD global head and the distilled local-descriptor decoder are concurrent consumers of the same DINOv2-B token field F_B, so descriptor training cannot change the retrieval descriptor. The shared token field replaces a second foundation-model backbone; only a 5.6M-parameter decoder is trained. The reference database stores pixel-aligned ECEF tensors, converting 2D matches directly into 2D-3D correspondences for PnP.

What would settle it

A decisive test: run RIM on a viewpoint-diverse dataset where query pitch varies from nadir to about 45° oblique, and compare its re-ranking Recall@1 against a version using the same distilled decoder but a separate, larger encoder (no sharing). If the shared-representation version falls more than ~1 percentage point behind, the single token field is the limiting factor; if it stays within 1 point, the sharing assumption holds. A second, cheaper check: perturb the reference mesh's ECEF coordinates by several meters and measure how much RIM's PnP inlier count drops—if the drop is large, the pix

Watch

Extended reading notes

Core claim

The central discovery is that the token field of a single DINOv2-B backbone, after a two-stage cross-domain fine-tuning on UAV-viewpoint renderings, contains enough information for both global place recognition and local feature matching. RIM reuses the full token field F_B for a frozen SALAD aggregation head and a distilled DeDoDe-style descriptor decoder, with the invariant g_after(I) = g_before(I) guaranteeing that local learning cannot corrupt retrieval. The distilled decoder, trained by knowledge distillation from a DINOv2-L teacher, produces a dense descriptor map that supports geometric re-ranking of Top-K candidates and, with pixel-aligned ECEF coordinates stored per reference pixel,

Load-bearing premise

The pipeline assumes that references rendered from a photogrammetric 3D mesh—sampled on a 20 m horizontal grid with discrete altitudes and orientations—are dense and geospatially accurate proxies for real UAV views, in appearance and in the pixel-aligned ECEF coordinates used for PnP; if the mesh is outdated, misaligned, or too sparse, both fine-tuning transfer and pose estimates degrade.

Editorial extensions

If this is right

  • A single DINOv2-B forward per query suffices for both retrieval and matching, so the query-side encoder cost is paid once regardless of re-ranking depth K.
  • The two-stage fine-tuning recipe—pose-near positives followed by offline-mined geographically distant hard negatives—lifts 3D Recall@1 over vanilla SALAD by 8.55/13.77 percentage points on Urbanscape and 4.45/8.94 on Chang'an Park at 25/50 m.
  • Geometric re-ranking using the distilled descriptors raises 3D R@1 by up to 7.28 percentage points and increases mean PnP inliers from 210 to 295 on Urbanscape.
  • At K=5 the end-to-end query runs in 67.9 ms, 1.8x faster than separate DeDoDe and 42.5x faster than RoMa, while remaining within ~1 point of the teacher's re-ranking Recall@1 and achieving 100% PnP success.
  • The pipeline reaches 93.00% PnP Recall@1 at 25 m on Urbanscape, enabling cold-start 6-DoF localization without an initial pose.

Reading between the lines

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

  • Because retrieval and matching share one backbone, the architecture could extend to other heads—semantic, depth, or object detection—on the same token field without additional encoder cost.
  • The single global +9.03 m vertical-datum correction applied to the Chang'an Park ground truth means the benchmark's absolute vertical accuracy is fixed by that one fitted offset; a per-flight or per-query re-estimation could change the reported Park pose numbers.
  • The offline hard-negative mining could be replaced by an online variant that re-encodes only the candidate pool, avoiding the risk that the Stage-2 pool under-represents hard negatives that appear only after further fine-tuning.
  • The 67.9 ms figure assumes a single RTX 3090 and 322x322 input; on lower-power flight hardware the relative latency gap between RIM and dense matchers is likely to widen, because dense matchers' per-candidate forward cost dominates.
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

2 major / 3 minor

Summary. The paper proposes RIM, a UAV global visual localization pipeline. Reference databases are built by rendering UAV-viewpoint views from Google Photorealistic 3D Tiles or a self-reconstructed SfM mesh, with pixel-aligned ECEF tensors stored per view. A two-stage cross-domain fine-tuning recipe adapts a SALAD/DINOv2-B retriever using pose-near positives and geographically distant hard negatives. RIM then freezes the adapted retriever and distils a DeDoDe-style local descriptor decoder on the same DINOv2-B token field, so one query-side backbone forward supplies both the global descriptor and the local descriptor map. Top-K candidates are re-ranked by geometric verification and the pose is estimated by RANSAC-PnP. The paper reports large Recall@1 gains over vanilla SALAD on two zero-shot benchmarks (EPFL Urbanscape and Chang'an Park) and end-to-end latency of 67.9 ms at K=5, about 1.8x faster than the separate DeDoDe sparse pipeline and over 40x faster than RoMa. The authors also provide matched-data retrained baselines, CUDA-synchronized latency measurements, a verified cost model, extensive ground-truth auditing for the Park benchmark, and explicit limitations.

Significance. If the headline claims hold, the paper makes a useful contribution: it demonstrates that a frozen DINOv2-B token field can simultaneously support global retrieval and local geometric matching, with Eq. (8) holding by construction rather than by empirical coincidence. The evaluation protocol is unusually careful in several respects: geographically disjoint test geographies are claimed, matched-data retrained baselines (NetVLAD*, SelaVPR++*) control for data access, the cost equations (10)-(11) are verified by the measured K-scaling in Appendix E, and the Park ground truth is audited in detail. The paper is also transparent about limitations, including mesh coverage, storage cost, and the nadir-geometry ceiling on Park. The main risk is not internal inconsistency but the current lack of support for the zero-shot disjointness claim: 'Urbanscape-OOP' is never defined or geolocated. Since the fine-tuning and test splits share the same rendering pipeline and both carry the 'Urbanscape' name, the zero-shot framing rests on a load-bearing assumption that the manuscript does not yet substantiate. A second, smaller risk is the sensitivity of the Park benchmark to its fitted vertical-datum co

major comments (2)
  1. [Abstract; §4.1; Table 1; Fig. B.2] The zero-shot claim depends on the assertion that the Urbanscape-OOP fine-tuning split is 'geographically disjoint from both test scenes' (Table 1 caption, §4.1). However, 'OOP' is never expanded, and the paper provides no coordinates, map, or overlap check for the fine-tuning and test reference grids. Both splits are rendered from Google Photorealistic 3D Tiles with the same CesiumJS sampler (§3.2), and both are called 'Urbanscape'. If OOP is an out-of-plane or sub-area subset of the same EPFL scene, the fine-tuning set may overlap the test reference grid, in which case the Table 2 gains would not demonstrate cross-domain generalization; the matched-data baselines would be contaminated in the same way. This is load-bearing for the central claim. Please define OOP precisely and provide coordinate-level evidence of disjointness (e.g., the distance between the convex hulls of the two refer
  2. [§4.1; Appendices I-J; Tables 3 and D.1] The Park benchmark's headline gains (+4.45/+8.94 pp in Table 3) are evaluated against PUBLIC v4 ground truth, which applies a global +9.03 m vertical-datum correction estimated from the STRICT subset and uses Oracle rotations on only 945 of the 2,315 queries, with EXIF fallback elsewhere. The 25/50 m thresholds are the same order of magnitude as the correction, so the absolute 3D Recall values could be sensitive to the fitting procedure. Although the audit is careful and relative method rankings are likely stable, the paper does not report how the main recall numbers change under a perturbed correction (e.g., within its 95% CI) or when evaluated only on the STRICT subset. Please add such a sensitivity analysis to confirm that the Park gains and deltas are not artifacts of the GT correction or the mixed provenance.
minor comments (3)
  1. [Tables 2, 3, 5, 6, 8] No confidence intervals or significance tests are reported for the head-to-head recall deltas. Many deltas are large, but the paper already uses Wilson bands for N=3158 in Appendix C; including CIs for the main tables would make the comparison more rigorous, especially for the +1.49 pp Stage-2 contribution.
  2. [§4.1 and Data availability] For reproducibility, please publish the exact geographic coordinates and sampling grid parameters of both the Urbanscape-OOP and test regions, as well as the script or metadata used to render the Google 3D Tiles views. Since the renderings themselves cannot be redistributed under Google Maps Platform terms, coordinate-level provenance is essential for verifying the disjointness claim.
  3. [Figure 4] The caption uses 'GMACs / TMACs' without defining the abbreviations; please spell out 'giga-multiply-accumulate' and 'tera-multiply-accumulate' at first use.

Circularity Check

1 steps flagged · score 1.0 of 10

No significant derivation-level circularity: RIM's central claims rest on frozen-architecture identities (Eq. 8), measured latencies (Table 7) and test-set recall (Tables 2-3). One concrete but non-central evaluation loop appears in the supplementary Park Oracle GT, where RoMa v2 is used to construct the rotation labels later used to score RoMa v2.

  1. other [Appendix J (Park PUBLIC ground truth) and Table D.1]
    "For every test query q we (i) take the nearest K=7 reference tiles by GT-position, (ii) run RoMa v2 dense 4K matching against each of them, (iii) solve a per-tile RANSAC PnP ... The PUBLIC v4 release combines ... (b) Oracle rotations on the same 945 STRICT queries ... Table D.1 lists 'RoMa v2' with 'R@5◦ 54.86'."

    The rotation labels used as ground truth in the Park evaluation are generated from RoMa v2's own dense matches, and a RoMa-based cross-query propagation pass extends them. Table D.1 then reports RoMa v2's rotation recall against those same labels, so the score is partly a function of the matcher being scored. This is an evaluation loop rather than a derivation loop: it is confined to the rotation columns of the supplementary Park ceiling analysis, does not involve the main Urbanscape/retrieval/latency claims, and the binding 5 m position channel in Table D.1 is RTK-derived rather than RoMa-derived.

full rationale

The paper's derivation chain is largely self-contained and empirical. Eq. (8), g_after(I) = g_before(I), is an identity that follows from freezing the adapted backbone and SALAD head during decoder distillation; it is presented as an architectural invariant, not as a derived prediction. Eqs. (10)-(11) are arithmetic cost decompositions whose two predicted behaviours are then verified by direct timing measurements in Section 4.5/Table 7 and Appendix E, not fitted to the reported 67.9 ms number. The two-stage fine-tuning gains are measured on test queries from EPFL/Urbanscape and Chang'an Park against released checkpoint baselines and matched-data retrained controls (Tables 2-3, L.1), so they are not predictions forced by the training objective. The 'Urbanscape-OOP' fine-tuning split is asserted to be geographically disjoint from both test scenes (Table 1 caption, Section 4.1) but the acronym is never expanded and no coordinates are given; this is a missing-support / potential leakage risk for the zero-shot claim, but it is not a definitional or fitted-input circularity, so it does not enter the score. The only concrete circular step is the Appendix J Oracle GT: RoMa v2 matching is used to build Park Oracle rotations (and RoMa warps propagate them), after which Table D.1 scores RoMa v2's rotation recall on that same GT. This loop is real but minor, non-central, and does not affect the headline retrieval, latency, Urbanscape pose, or the RTK-derived Park position bottleneck; hence the low score.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claims rest on rendered-reference validity, transfer of SALAD's GSV-Cities initialization, and the trustworthiness of two benchmark GTs. Most load-bearing are the 3D-Tiles/SfM rendering premise (Section 3.2) and the Park GT construction (Appendices I–J). No new physical entities are introduced; the free parameters are neural-net hyperparameters, mining thresholds, dataset-construction choices, and one fitted GT correction.

free parameters (6)
  • λd = 0.5 (Stage-1 self-distillation weight) = 0.5
    Eq. (3); chosen by hand; anchors fine-tuned SALAD descriptors to the original weights.
  • Stage-2 triplet margin = 0.3
    Section 4.2; margin for the triplet loss on mined geographically distant hard negatives.
  • Distillation loss weights (λmse, λcos, λsgn) and 20% hard-pixel fraction = (1.0, 0.5, 0.1), 20%
    Eq. (9); chosen by hand for the teacher–student local-descriptor objective.
  • Hard-negative mining thresholds (Top-M, τd, τs) = yields <100 strict triplets
    Algorithm A.1; thresholds chosen by hand; tiny pool compensated by 10× oversampling.
  • Park GT vertical-datum correction = +9.03 m
    Appendix I; median Oracle−IMU vertical residual on 945 STRICT queries, applied globally to all Park query altitudes; directly affects the 3D recall labels.
  • Reference grid sampling (horizontal step, altitude/pitch/yaw sets) = 20 m; {500–800 m}, −75°; {530–570 m}, {−90°, −60°}, 45° yaw steps
    Section 3.2/4.1; dataset-construction choices that set reference density and therefore the achievable localization tolerance.
assumptions (5)
  • domain assumption Rendered UAV-perspective views from Google 3D Tiles / SfM meshes are an adequate appearance and geometry proxy for real UAV imagery after fine-tuning.
    Section 3.2; the entire database-construction mechanism; acknowledged as coverage-limited in Section 5.
  • domain assumption A single DINOv2-B token field contains both retrieval-salient and local-matching-salient information.
    Section 3.4; the premise of RIM — the distilled decoder learns to interpret the frozen retrieval tokens; validated empirically, not theoretically.
  • domain assumption SALAD's GSV-Cities-trained initialization transfers to the rendered-UAV domain.
    Section 3.3; Stage 1 starts from GSV-Cities SALAD; Appendix C shows dropping the SALAD initialization collapses R@1 to 39.3%.
  • domain assumption The Chang'an Park 'PUBLIC v4' ground truth is accurate within the evaluation tolerances.
    Appendices I–J; hybrid RTK + corrected-IMU + Oracle construction; the +9.03 m correction and σz≈8 m residual are comparable in size to the 25 m tolerance.
  • standard math RANSAC-PnP / fundamental-matrix estimation converge to the correct pose given the candidate matches.
    Eq. (12); standard geometric tools used off-the-shelf with USAC-MAGSAC.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RIM: A Retrieval-In-Matching Framework for Cross-Domain Global Visual Localization of UAVs." pith.science (2026). https://pith.science/paper/I5PASVAP

@misc{pith2026260720116,
  author       = {Pith},
  title        = {Pith review of: RIM: A Retrieval-In-Matching Framework for Cross-Domain Global Visual Localization of UAVs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I5PASVAP}},
  note         = {Machine review of arXiv:2607.20116}
}
read the original abstract

Global visual localization of unmanned aerial vehicles (UAVs) using remote-sensing reference maps has attracted increasing attention. However, acquisition-time and imaging-platform differences between UAV and reference imagery induce substantial cross-domain appearance and viewpoint shifts, challenging robust six-degree-of-freedom (6-DoF) pose estimation. We address these shifts by sampling UAV-viewpoint reference views from Google 3D Tiles across locations, altitudes, and orientations. A two-stage cross-domain fine-tuning recipe adapts SALAD using pose-near positives and geographically distant hard negatives, while local geometric consistency re-ranks the Top-K candidates. We further propose Retrieval-In-Matching (RIM), which freezes the adapted DINOv2-B retriever and distils a local-descriptor decoder that reuses its token field alongside a shallow VGG19 detail stream. One query-side DINOv2-B forward thus serves both SALAD retrieval and local description, eliminating a second foundation-model backbone while preserving retrieval descriptors by construction. We evaluate RIM zero-shot on the reconstructed EPFL Urbanscape and self-collected Chang'an Park datasets, both geographically disjoint from the training data. RIM outperforms ten recent retrieval baseline families. At 25/50 m under the full 3D distance metric, it improves Recall@1 over SALAD by 8.55/13.77 percentage points on EPFL and 4.45/8.94 points on Park. At Top-K=5, the complete measured localization query, including retrieval, candidate matching, and robust geometric verification, takes 67.9 ms end-to-end: 1.8 times faster than the strongest separate sparse-matching baseline and over 40 times faster than RoMa, while achieving comparable re-ranking accuracy. These results establish an efficient and deployable pipeline for UAV global visual localization in GNSS-challenged environments.

Figures

Figures reproduced from arXiv: 2607.20116 by the authors.

Figure 1
Figure 1. UAV-viewpoint reference sampling and two-stage cross-domain retriever adap [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Overview of the Retrieval-In-Matching (RIM) framework. [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Cross-domain benchmark overview: two representative real UAV queries (black [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Accuracy–speed–size Pareto view on Urbanscape. Each marker is one [PITH_FULL_IMAGE:figures/full_fig_p025_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 2 canonical work pages

  1. [1]

    T. Qin, P. Li, S. Shen, VINS-Mono: A robust and versatile monocular visual-inertial state estimator, IEEE Transactions on Robotics 34 (2018) 1004–1020. doi:10.1109/TRO.2018.2853729

    arXiv 2018

  2. [2]

    Campos, R

    C. Campos, R. Elvira, J. J. Gómez Rodríguez, J. M. M. Montiel, J. D. Tardós, ORB-SLAM3: An accurate open-source library for visual, visual-inertial, and multimap SLAM, IEEE Transactions on Robotics 37 (2021) 1874–1890. doi:10.1109/TRO.2021.3075644. 28

    arXiv 2021

  3. [3]

    J. Zhu, S. Peng, L. Wang, H. Tan, Y. Liu, M. Zhang, S. Yan, LoD- Loc v2: Aerial visual localization over low level-of-detail city models using explicit silhouette alignment, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp. 26610– 26621

    2025

  4. [4]

    Izquierdo, J

    S. Izquierdo, J. Civera, Optimal transport aggregation for visual place recognition, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 17658–17668

    2024

  5. [5]

    Ali-Bey, B

    A. Ali-Bey, B. Chaib-Draa, P. Giguère, GSV-Cities: Toward appropriate supervised visual place recognition, Neurocomputing 513 (2022) 194–

    2022

  6. [6]

    Deuser, K

    F. Deuser, K. Habel, N. Oswald, Sample4Geo: Hard negative sam- pling for cross-view geo-localisation, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 16847– 16856

    2023

  7. [7]

    URL: https://openreview.net/forum?id=a68SUt6zFt

    M.Oquab, T.Darcet, T.Moutakanni, H.V.Vo, M.Szafraniec, V.Khali- dov, et al., DINOv2: Learning robust visual features without su- pervision, Transactions on Machine Learning Research (2024). URL: https://openreview.net/forum?id=a68SUt6zFt

    2024

  8. [8]

    Edstedt, G

    J. Edstedt, G. Bökman, M. Wadenbäck, M. Felsberg, DeDoDe: Detect, don’t describe – describe, don’t detect for local feature matching, in: Proceedings of the International Conference on 3D Vision (3DV), 2024, pp. 148–157. doi:10.1109/3DV62453.2024.00035

    arXiv 2024

Show all 46 references
  1. [9]

    Goforth, S

    H. Goforth, S. Lucey, GPS-denied UAV localization using pre-existing satellite imagery, in: International Conference on Robotics and Automa- tion, ICRA 2019, Montreal, QC, Canada, May 20–24, 2019, IEEE, 2019, pp. 2974–2980. URL:https://doi.org/10.1109/ICRA.2019.8793558. doi:10...

    2019

  2. [10]

    Bianchi, T

    M. Bianchi, T. D. Barfoot, UAV localization using autoencoded satellite images, IEEE Robotics and Automation Letters 6 (2021) 1761–1768. doi:10.1109/LRA.2021.3060397

    2021

  3. [11]

    S. Ahn, H. Kang, J. Lee, Aerial-satellite image matching framework for UAV absolute visual localization using contrastive learning, in: 2021 29 21st International Conference on Control, Automation and Systems (IC- CAS), 2021, pp. 143–146. doi:10.23919/ICCAS52745.2021.9649999

    2021

  4. [12]

    H. Sui, J. Li, J. Lei, C. Liu, G. Gou, A fast and robust heterologous image matching method for visual geo-localization of low-altitude UAVs, Remote Sensing 14 (2022) 5879. doi:10.3390/rs14225879

    2022 doi

  5. [13]

    Zheng, Y

    Z. Zheng, Y. Wei, Y. Yang, University-1652: A multi-view multi-source benchmark for drone-based geo-localization, in: Proceedings of the 28th ACM International Conference on Multimedia, 2020, pp. 1395–1403

    2020

  6. [14]

    R. Wu, J. Deng, M. Mou, X. He, M. Zhang, Y. Liu, S. Yan, UAV- GeoLoc: A large-vocabulary dataset and geometry-transformed method for UAV geo-localization, IEEE Robotics and Automation Letters 10 (2025) 8978–8985. doi:10.1109/LRA.2025.3588061

    2025

  7. [15]

    Moskalenko, A

    I. Moskalenko, A. Kornilova, G. Ferrer, Visual place recognition for aerial imagery: A survey, Robotics and Autonomous Systems 183 (2025) 104837. doi:10.1016/j.robot.2024.104837

    2025

  8. [16]

    W. Xu, Y. Yao, J. Cao, Z. Wei, C. Liu, J. Wang, M. Peng, UAV- VisLoc: Alarge-scaledatasetforUAVvisuallocalization, arXivpreprint arXiv:2405.11936 (2024)

    2024 arXiv

  9. [17]

    Y. Ji, B. He, Z. Tan, L. Wu, Game4Loc: A UAV geo-localization benchmark from game data, in: Proceedings of the AAAI Con- ference on Artificial Intelligence, volume 39, 2025, pp. 3913–3921. doi:10.1609/aaai.v39i4.32409

    2025 doi

  10. [18]

    Y. Ji, B. He, Z. Tan, L. Wu, MMGeo: Multimodal compositional geo- localization for UAVs, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp. 25165–25175

    2025

  11. [19]

    Q. Yan, J. Zheng, S. Reding, S. Li, I. Doytchinov, CrossLoc: Scalable aerial localization assisted by multimodal synthetic data, in: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 17358–17368

    2022

  12. [20]

    H. Li, J. Wang, Z. Wei, W. Xu, Jointly optimized global-local visual localization of UAVs, arXiv preprint arXiv:2310.08082 (2023). 30

    2023 arXiv

  13. [21]

    Arandjelovic, P

    R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, J. Sivic, NetVLAD: CNN architecture for weakly supervised place recognition, in: Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2016, pp. 5297–5307

    2016

  14. [22]

    H. Noh, A. Araujo, J. Sim, T. Weyand, B. Han, Large-scale image retrieval with attentive deep local features, in: Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 3456– 3465

    2017

  15. [23]

    Radenović, G

    F. Radenović, G. Tolias, O. Chum, Fine-tuning CNN image retrieval with no human annotation, IEEE Transactions on Pattern Analysis and Machine Intelligence 41 (2019) 1655–1668. doi:10.1109/TPAMI.2018. 2846566

    2019 doi

  16. [24]

    Berton, C

    G. Berton, C. Masone, B. Caputo, Rethinking visual geo-localization for large-scale applications, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 4878– 4888

    2022

  17. [25]

    Ali-Bey, B

    A. Ali-Bey, B. Chaib-Draa, P. Giguère, MixVPR: Feature mixing for visual place recognition, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023, pp. 2998–3007

    2023

  18. [26]

    Berton, G

    G. Berton, G. Trivigno, B. Caputo, C. Masone, EigenPlaces: Training viewpoint robust models for visual place recognition, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 11080–11090

    2023

  19. [27]

    Ali-Bey, B

    A. Ali-Bey, B. Chaib-Draa, P. Giguère, BoQ: A place is worth a bag of learnable queries, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 17794– 17803

    2024

  20. [28]

    Keetha, A

    N. Keetha, A. Mishra, J. Karhade, K. M. Jatavallabhula, S. Scherer, M. Krishna, S. Garg, AnyLoc: Towards universal visual place recog- nition, IEEE Robotics and Automation Letters 9 (2024) 1286–1293. doi:10.1109/LRA.2023.3343602. 31

    2024

  21. [29]

    F. Lu, T. Jin, X. Lan, L. Zhang, Y. Liu, Y. Wang, C. Yuan, Sela- VPR++: Towardsseamlessadaptationoffoundationmodelsforefficient place recognition, IEEE Transactions on Pattern Analysis and Machine Intelligence (2025). doi:10.1109/TPAMI.2025.3629287, early Access

    2025

  22. [30]

    DeTone, T

    D. DeTone, T. Malisiewicz, A. Rabinovich, SuperPoint: Self-supervised interest point detection and description, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018, pp. 224–236

    2018

  23. [31]

    Sarlin, D

    P.-E. Sarlin, D. DeTone, T. Malisiewicz, A. Rabinovich, SuperGlue: Learning feature matching with graph neural networks, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), 2020, pp. 4938–4947

    2020

  24. [32]

    J. Sun, Z. Shen, Y. Wang, H. Bao, X. Zhou, LoFTR: Detector-free local feature matching with transformers, in: Proceedings of the IEEE/CVF ConferenceonComputerVisionandPatternRecognition(CVPR),2021, pp. 8922–8931

    2021

  25. [33]

    Edstedt, I

    J. Edstedt, I. Athanasiadis, M. Wadenbäck, M. Felsberg, DKM: Dense kernelized feature matching for geometry estimation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2023, pp. 17765–17775

    2023

  26. [34]

    Edstedt, Q

    J. Edstedt, Q. Sun, G. Bökman, M. Wadenbäck, M. Felsberg, RoMa: Robust dense feature matching, in: Proceedings of the IEEE/CVF ConferenceonComputerVisionandPatternRecognition(CVPR),2024, pp. 19790–19800

    2024

  27. [35]

    X. Zhao, X. Wu, W. Chen, P. C. Y. Chen, Q. Xu, Z. Li, ALIKED: A lighter keypoint and descriptor extraction network via deformable trans- formation, IEEE Transactions on Instrumentation and Measurement 72 (2023) 1–16. doi:10.1109/TIM.2023.3271000

    2023

  28. [36]

    Y. Wang, X. He, S. Peng, D. Tan, X. Zhou, Efficient LoFTR: Semi-dense local feature matching with sparse-like speed, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 21666–21675. 32

    2024

  29. [37]

    X. He, H. Yu, S. Peng, D. Tan, Z. Shen, H. Bao, X. Zhou, MatchAny- thing: Universal cross-modality image matching with large-scale pre- training, arXiv preprint arXiv:2501.07556 (2025)

    2025 arXiv

  30. [38]

    Accessed: 2026-07-22

    Cesium GS, Inc., CesiumJS: An open-source JavaScript library for world-class 3D globes and maps,https://cesium.com/platform/ cesiumjs/, n.d. Accessed: 2026-07-22

    2026

  31. [39]

    Google LLC, Google photorealistic 3D tiles: Map tiles API,https: //developers.google.com/maps/documentation/tile/3d-tiles,

  32. [40]

    X. Wang, X. Han, W. Huang, D. Dong, M. R. Scott, Multi-similarity loss with general pair weighting for deep metric learning, in: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 5022–5030

    2019

  33. [41]

    Sarlin, C

    P.-E. Sarlin, C. Cadena, R. Siegwart, M. Dymczyk, From coarse to fine: Robust hierarchical localization at large scale, in: Proc. IEEE/CVF ConferenceonComputerVisionandPatternRecognition(CVPR),2019, pp. 12716–12725

    2019

  34. [42]

    Revaud, J

    J. Revaud, J. Almazan, R. S. Rezende, C. R. de Souza, Learning with average precision: Training image retrieval with a listwise loss, in: Proc. IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 5107–5116

    2019

  35. [43]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, I. Sutskever, Learning transferable visual models from natural language supervision, in: Proceedings of the 38th International Conference on Machine Learn-...

    2021

  36. [44]

    before re-rank

    P. Lindenberger, P.-E. Sarlin, M. Pollefeys, LightGlue: Local feature matching at light speed, in: Proc. IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 17627–17638. 33 Supplementary Material Appendix A. Offline Hard-Negative Mining Procedure The main pa...

    2023

  37. [203]

    doi:10.1016/j.neucom.2022.09.127

    2022 doi

  38. [2026]

    Accessed: 2026-07-22

    2026

Pith tools

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