Pith. sign in

REVIEW 4 major objections 6 minor 49 references

This paper claims that DINO-ViT semantic features plus coarse-to-fine FFT correlation localize a single image in OpenStreetMap faster and more accurately than prior methods, with 3° heading recall beating rivals' 5°.

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

T0 review · deepseek-v4-flash

2026-08-02 19:33 UTC pith:J23FCDKT

load-bearing objection Solid engineering paper on OSM-based localization with a new DINO-BEV + FFT coarse-to-fine pipeline; the KITTI comparison is not apples-to-apples because of an unacknowledged ±10° heading prior. the 4 major comments →

arxiv 2603.01613 v2 pith:J23FCDKT submitted 2026-03-02 cs.CV

Uncertainty-Aware Hierarchical Re-Localization in OpenStreetMap via Semantic Alignment

classification cs.CV MSC 68T45
keywords monocular re-localizationOpenStreetMapDINO-ViTsemantic alignmentcoarse-to-fine searchFFT correlationbird's-eye viewpose estimation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that a single monocular image can be localized against OpenStreetMap vector data both more accurately and about three times faster than the previous best method. The key move is to use DINO-ViT's self-supervised semantic tokens to project the image into a bird's-eye-view semantic feature map that can be matched to a rasterized OSM tile, and then to replace exhaustive dense matching with a coarse-to-fine search: a cheap FFT correlation on low-resolution features gives an initial pose, and an uncertainty measure derived from the pose cost volume adaptively shrinks the search window for a fine 0.5-meter, 2-degree refinement. Results on MGL and KITTI show the 3-degree heading recall beating the prior method's 5-degree recall, with FPS rising from about 7.5 to about 26.

Core claim

On the paper's terms, the central discovery is that object-centric DINO-ViT tokens, projected to BEV through a depth-based polar-to-Cartesian resampling, align ground-view photographs with OpenStreetMap's vector semantics well enough that a coarse-to-fine FFT correlation search recovers 3-DoF pose at high recall. The paper further claims that the uncertainty-controlled refinement—shrinking the angular and spatial search ranges according to a softmax over neighboring pose hypotheses—is what converts the coarse pose into fine pose while keeping speed high.

What carries the argument

The load-bearing identity is the Fourier-domain correlation M(p,θ) = F^{-1}(F(F_map) ⊙ F(F_bev(θ))) over a rotated semantic BEV, which turns dense cross-correlation into element-wise multiplication. DINOv2 tokens (a self-supervised vision transformer with registers, frozen in the first eight layers and fine-tuned in the last four) supply the semantic features; the uncertainty measure U = 1 − exp(M(p_c,θ_c)) / Σ_{(p,θ)∈N} exp(M(p,θ)) then sets the fine search window size. Together these make the coarse-to-fine hierarchy work end-to-end.

Load-bearing premise

The true pose must lie inside the 128×128 meter OSM tile centered on the query's GPS position; if the GPS error exceeds that tile, the method cannot recover the pose.

What would settle it

Take the MGL test set and artificially shift the provided GPS positions by 80 meters (outside the tile) or remove the GPS anchor entirely; if recall drops to near zero, the method is confirmed to be tile-anchored refinement rather than global localization. Similarly, run KITTI without the ±10° heading prior and measure whether 3° recall stays above 80%.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • On MGL, fine localization reaches 17.36% recall at 1 meter, 53.67% at 3 meters, and 65.81% at 5 meters; orientation recall reaches 30.66% at 1°, 68.71% at 3°, and 80.53% at 5°.
  • On KITTI trained from scratch, the method's 3° heading recall (80.16%) exceeds OrienterNet's 5° recall (73.53%), and 3m position recall (92.84%) exceeds the prior 5m recall (91.81%).
  • The coarse stage alone runs at 29.54 FPS and fine at 25.80 FPS on MGL, versus 7.51 FPS for OrienterNet, a more than threefold speedup.
  • Ablations attribute most of the accuracy gain to DINO features and most of the speed gain to the coarse-to-fine strategy; adding DINO alone improves recalls but lowers FPS slightly.

Where Pith is reading between the lines

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

  • If the GPS anchor were replaced by a global retrieval stage that proposes candidate tiles, the same coarse-to-fine pipeline could in principle perform true global localization instead of refinement within a 128×128 m tile.
  • The uncertainty-based window shrinking is a generic mechanism: it could be applied to other cross-modal matching tasks (satellite-to-ground, radar-to-map) wherever a pose cost volume is available.
  • One testable extension is to quantify how the 3°-over-5° advantage scales with smaller angular step sizes; the paper's 2° fine step suggests heading accuracy is limited by discretization more than by feature quality.
  • Because DINO's zero-shot semantics help most on the small KITTI training set, the method may generalize to cities and sensors unseen at training time, which is testable by evaluating on MGL cities held out from fine-tuning.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a coarse-to-fine monocular re-localization method that matches a query image to OpenStreetMap (OSM). The query image is processed by DINOv2 to obtain semantic tokens, which are projected into a bird's-eye-view (BEV) feature map; the OSM tile is rasterized and embedded into a neural semantic map. Global dense matching is replaced by a coarse FFT-correlation search over position and orientation, followed by a fine stage whose search range is contracted using an uncertainty measure derived from the coarse pose likelihood. Experiments on MGL and KITTI report higher position/orientation recall and roughly 3× higher FPS relative to OrienterNet, with the abstract claiming that a 3° orientation recall outperforms the 5° recall of prior state of the art.

Significance. If the results are robust, the paper makes a practical contribution: it shows that OSM-only, privacy-preserving localization can be made substantially faster and more accurate than the existing OrienterNet baseline, and it provides a clean use of DINO semantic features to bridge the image-to-map modality gap. The coarse-to-fine FFT matching is technically sound and the speedup is plausible from the reduced search space. However, the evaluation protocol on KITTI gives the proposed method a ±10° orientation prior that baselines do not receive, and the MGL protocol uses a GPS-anchored 128×128 m tile despite claiming 'no pose prior.' These issues matter because the headline orientation-recall comparison is directly affected. The core idea is still defensible, but the claims need to be re-scoped and the comparison made fair before the results can be accepted as stated.

major comments (4)
  1. [V-B / Table II] The KITTI comparison is confounded by an asymmetric orientation prior. Section V-B states that for KITTI 'sample points are selected within a ±10° range of initial pose,' whereas OrienterNet is evaluated over the full 360° orientation range in its original protocol. This directly inflates the reported orientation recall: the headline '3° recall outperforms 5° recall' (Abstract) relies on the KITTI row where Ours-KITTI 3° is 80.16 vs OrienterNet-KITTI 5° 73.53. Please re-evaluate OrienterNet under the same ±10° prior, or report your method without that prior, and remove or condition the superiority claim on the prior. The finer 2° angular grid (Section V-A) also pre-orders comparisons at 3° vs 5°; this should be quantified by evaluating both methods at a common set of thresholds.
  2. [V-A / V-B] The statement 'we do not use pose prior' for MGL is misleading. Section V-A says OSM tiles are 'acquired from the noisy GPS position information attached to the MGL image,' and the tile is 128×128 m. This is a spatial prior: the true pose is assumed to lie inside a GPS-centered tile. If GPS error exceeds roughly half the tile size, the method cannot recover. The method is therefore local refinement within a GPS-anchored tile, not global localization. Please clarify whether this is the same protocol as OrienterNet and evaluate sensitivity to GPS noise, e.g., by shifting the tile center by increasing offsets and reporting recall.
  3. [IV / V-B] The training objective is never specified. Section II-B promises 'end-to-end training through pose supervision,' but neither the loss function nor the training schedule is given. Is the network trained with a cross-entropy loss over the pose cost volume of Eq. (4), or with a regression loss on M(p,θ)? What is the optimizer, learning rate, batch size, and number of epochs? Without this information the method is not reproducible and it is not possible to judge whether Eq. (4) is actually a learned likelihood or just a handcrafted correlation. Please specify the full training protocol and any losses used for the DINO fine-tuning and the map encoder.
  4. [IV-C, Eqs. (6)-(8)] The uncertainty measure U is a central component, but its calibration is not validated. The paper does not show that U correlates with pose error, nor does it compare the uncertainty-contracted refinement against a fixed-width refinement window. If U is poorly calibrated, the contraction in Eqs. (7)-(8) could reduce recall instead of improving it. Please add an analysis of U (e.g., reliability diagrams or scatter of U vs. error) and an ablation that uses a fixed search range of comparable average size, to demonstrate that the adaptive contraction, rather than the extra fine stage alone, is responsible for the reported gains.
minor comments (6)
  1. [Title/Abstract] The arXiv metadata title 'Uncertainty-Aware Hierarchical Re-Localization...' differs from the full-text title 'Coarse-to-Fine Monocular Re-Localization...'. The two should be aligned.
  2. [Tables I–II] No error bars or statistical significance tests are reported. Given the claim of 'significant improvements' and the stochasticity in training, please report mean±std over at least 3 runs, or at least state the number of runs.
  3. [Table I] FPS is reported without specifying the measurement protocol (GPU model, batch size, whether preprocessing is included, whether OrienterNet was re-run on the same hardware or taken from its paper). Please clarify so the speed comparison is fair.
  4. [V-A] The MGL dataset description says 'approximately 760,000 images' but does not state the number used for training/validation/test, nor how the split is made. Please give dataset statistics and the exact split used.
  5. [III] The PCA visualization in Fig. 2 is qualitative. A quantitative comparison of the semantic alignment ability of DINO vs. VGG/ResNet (e.g., segmentation IoU on a few categories) would strengthen the motivation.
  6. [IV-B] The channel dimension N is said to be aligned with F_bev but its numerical value is not given. Also, '7 categories for areas, 10 for roads, 33 for points' is stated without explaining whether all are used in the semantic embedding and whether some are discarded during training.

Circularity Check

0 steps flagged

No circular dependence: method is trained on held-out supervision and evaluated on MGL/KITTI test sets.

full rationale

The paper's prediction pipeline is not circular. The method uses DINOv2 features, BEV projection, an OSM-derived semantic map, and coarse-to-fine FFT correlation; pose supervision is provided by ground-truth poses in MGL and KITTI, and evaluation is on held-out test splits. The coarse-to-fine mechanism (Eqs. 3-8) uses the matching likelihood, not the ground-truth pose or headline metric, so the reported accuracy is not forced by construction. The only notable concerns are methodological fairness issues, not circularity: (i) KITTI testing uses an explicit ±10° heading prior (Section V-B: 'For KITTI dataset, we incorporate prior pose during testing, with sample points selected within a ±10° range of initial pose'), which weakens the apples-to-apples comparison with full-360° OrienterNet, and (ii) OSM tiles are centered on noisy GPS ('OSM tiles are acquired from the noisy GPS position information attached to the MGL image'), so the method performs GPS-anchored local refinement rather than global localization. These are scope/fairness limitations. The finer 2° angular grid versus OrienterNet's 5.625° may explain part of the orientation-recall advantage, but that is a resolution choice, not a predefined output. No load-bearing self-citations, imported uniqueness theorems, or fitted-parameter-renamed-as-prediction steps were found.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method's success rests on unproven modeling assumptions (DINO semantics align with OSM categories; BEV projection via depth distributions is reliable; a GPS-centered tile contains the pose) and on several hand-chosen parameters that are unreported or under-ablated. No new physical entities are introduced.

free parameters (5)
  • Coarse/fine angular search intervals = 6° coarse, 2° fine
    Manual search-grid choices; the finer fine-grid is the proximate cause of the 3°-vs-5° orientation comparison.
  • Coarse/fine map resolutions = 1 m coarse, 0.5 m fine
    Hand-chosen resolutions; not derived, and directly affect reported position recall.
  • Uncertainty contraction coefficients k_θ and k_p = not reported
    Control the refinement window in Eqs. (7)-(8) and are never given values, so the refinement behavior is unreproducible.
  • DINO layer freeze/fine-tune split = freeze first 8, fine-tune last 4
    Manual choice to reduce overfitting; not ablated or justified quantitatively.
  • Training sample radius and KITTI test prior = 32 m training radius; ±10° KITTI test prior
    These bounds define the difficulty of the search; the KITTI comparison depends on the ±10° prior.
axioms (5)
  • domain assumption DINO tokens encode semantic classes such as buildings, roads, and poles consistently across street views and OSM rendering.
    Section IV-A relies on DINO to decouple semantic elements; Fig. 2 provides only qualitative PCA visualizations, no quantitative validation.
  • domain assumption Ground-view features can be projected into BEV using a depth-based interpolation determined by camera intrinsics.
    Section IV-A, Eqs. (1)-(2); the depth distribution α(d,v) is not specified and is inherited from [14]. Wrong depths break the semantic alignment.
  • domain assumption The GPS-centered 128×128 m OSM tile contains the true pose.
    Section V-A acquires tiles from noisy GPS; Section V-B claims 'no pose prior.' Localization is thus bounded by GPS accuracy.
  • ad hoc to paper The FFT correlation response in Eq. (4) is a usable pose likelihood and the uncertainty measure U in Eq. (6) is calibrated.
    Eqs. (4)-(6) are introduced without calibration or ablation studies; if U is miscalibrated, the refinement range is wrong.
  • domain assumption MapLocNet is not a viable baseline for monocular MGL because the authors' unreported single-view training attempt failed.
    Section VI states 'Our attempt to use single-view training did not yield satisfactory results' with no details, excluding the closest coarse-to-fine OSM baseline.

pith-pipeline@v1.3.0-alltime-deepseek · 10572 in / 15578 out tokens · 135560 ms · 2026-08-02T19:33:23.021009+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Uncertainty-Aware Hierarchical Re-Localization in OpenStreetMap via Semantic Alignment." pith.science (2026). https://pith.science/paper/J23FCDKT

@misc{pith2026260301613,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-Aware Hierarchical Re-Localization in OpenStreetMap via Semantic Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J23FCDKT}},
  note         = {Machine review of arXiv:2603.01613}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Monocular re-localization enables robots to estimate camera poses from visual observations. However, many existing methods rely on dense maps or large reference image databases, which face scalability limitations and privacy risks. OpenStreetMap (OSM), as a lightweight privacy-preserving map, offers semantic and geometric information with global scalability. Nonetheless, OSM localization remains challenging due to cross-modal discrepancies between natural images and OSM, as well as the high cost of global map-based localization. In this paper, we propose an uncertainty-aware hierarchical search framework with semantic alignment for localization in OSM. First, object-centric DINO-ViT tokens are exploited to reduce the semantic gap between ground-view observations and OSM vectors. Second, global dense matching is decomposed into coarse FFT correlation and uncertainty-controlled local refinement. Extensive experiments demonstrate that our method significantly improves localization accuracy and speed. When trained on a single dataset, the 3$^\circ$ orientation recall of our method even outperforms the 5$^\circ$ recall of state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2603.01613 by Lihuang Fang, Xiao Hu, Yuchen Zou, Yuqing Tang.

Figure 1
Figure 1. Figure 1: Localization in the OSM-derived semantic map by transforming [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of street view image features from VGG19, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: First, we use DINO to extract the semantic features [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: Coarse-to-fine semantic localization framework for OSM localization. First, the query image is transformed into semantic features. Then, the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of the coarse-to-fine localization process, including refinement of the neural map transformed from OSM, and refinement of the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

49 extracted references · 1 linked inside Pith

  1. [1]

    Extracting dynamic navigation goal from natural language dialogue,

    L. Liang, G. Bian, H. Zhao, Y . Dong, and H. Liu, “Extracting dynamic navigation goal from natural language dialogue,” inIROS, 2023, pp. 3539–3545

  2. [2]

    Loc-nerf: Monte carlo localization using neural radiance fields,

    D. Maggio, M. Abate, J. Shi, C. Mario, and L. Carlone, “Loc-nerf: Monte carlo localization using neural radiance fields,” inICRA. IEEE, 2023, pp. 4018–4025

  3. [3]

    Tm 3loc: Tightly-coupled monocular map matching for high precision vehicle localization,

    T. Wen, K. Jiang, B. Wijaya, H. Li, M. Yang, and D. Yang, “Tm 3loc: Tightly-coupled monocular map matching for high precision vehicle localization,”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 11, pp. 20 268–20 281, 2022

  4. [4]

    Determining the impact of augmented reality graphic spatial location and motion on driver behaviors,

    M. Smith, J. L. Gabbard, G. Burnett, C. Hare, H. Singh, and L. Skrypchuk, “Determining the impact of augmented reality graphic spatial location and motion on driver behaviors,”Applied ergonomics, vol. 96, p. 103510, 2021

  5. [5]

    Representing remote locations with location-based augmented reality game design,

    Y . S. W. W. Arachchige, H. K. Lukosch, J. Everett, and S. Lukosch, “Representing remote locations with location-based augmented reality game design,”Entertainment Computing, p. 100932, 2025

  6. [6]

    A survey on monocular re- localization: From the perspective of scene map representation,

    J. Miao, K. Jiang, T. Wen, Y . Wang, P. Jia, B. Wijaya, X. Zhao, Q. Cheng, Z. Xiao, J. Huang,et al., “A survey on monocular re- localization: From the perspective of scene map representation,”IEEE Transactions on Intelligent V ehicles, 2024

  7. [7]

    Text2loc: 3d point cloud localization from natural language,

    Y . Xia, L. Shi, Z. Ding, J. F. Henriques, and D. Cremers, “Text2loc: 3d point cloud localization from natural language,” inCVPR. IEEE, 2024, pp. 14 958–14 967

  8. [8]

    Where is your place, visual place recognition?

    S. Garg, T. Fischer, and M. Milford, “Where is your place, visual place recognition?” inIJCAI, 2021, pp. 4416–4425

  9. [9]

    The revisiting problem in simultaneous localization and mapping: A survey on visual loop closure detection,

    K. A. Tsintotas, L. Bampis, and A. Gasteratos, “The revisiting problem in simultaneous localization and mapping: A survey on visual loop closure detection,”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 11, pp. 19 929–19 953, 2022

  10. [10]

    High-definition maps: Comprehensive survey, challenges, and future perspectives,

    G. Elghazaly, R. Frank, S. Harvey, and S. Safko, “High-definition maps: Comprehensive survey, challenges, and future perspectives,” IEEE Open Journal of Intelligent Transportation Systems, vol. 4, pp. 527–550, 2023

  11. [11]

    Lod-loc: Aerial visual localization using lod 3d map with neural wireframe alignment,

    J. Zhu, S. Yan, L. Wang, S. Zhang, Y . Liu, and M. Zhang, “Lod-loc: Aerial visual localization using lod 3d map with neural wireframe alignment,” inNeurIPS, 2024

  12. [12]

    Osmstats - statistics of the free wiki world map

    OSMstats, “Osmstats - statistics of the free wiki world map.” [Online]. Available: https://osmstats.neis-one.org/

  13. [13]

    Map it anywhere (mia): Empowering bird’s eye view mapping using large-scale public data,

    C. Ho, J. Zou, O. Alama, S. M. J. Kumar, B. Chiang, T. Gupta, C. Wang, N. Keetha, K. Sycara, and S. Scherer, “Map it anywhere (mia): Empowering bird’s eye view mapping using large-scale public data,” inAdvances in Neural Information Processing Systems, 2024

  14. [14]

    Ori- enternet: Visual localization in 2d public maps with neural matching,

    P. Sarlin, D. DeTone, T. Yang, A. Avetisyan, J. Straub, T. Malisiewicz, S. R. Bul `o, R. A. Newcombe, P. Kontschieder, and V . Balntas, “Ori- enternet: Visual localization in 2d public maps with neural matching,” inCVPR, 2023, pp. 21 632–21 642

  15. [15]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. As- sran, N. Ballas, W. Galuba, R. Howes, P. Huang, S. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. J ´egou, J. Mairal, P. Labatut, A. Joulin, and P. Bojanowski, “Dinov2: Learning robust visual features without sup...

  16. [16]

    U-BEV: height-aware bird’s-eye-view segmentation and neural map- based relocalization,

    A. B. Camiletto, A. Bochicchio, A. Liniger, D. Dai, and A. Gawel, “U-BEV: height-aware bird’s-eye-view segmentation and neural map- based relocalization,” inIROS. IEEE, 2024, pp. 5597–5604

  17. [17]

    A survey on visual-based localization: On the benefit of heterogeneous data,

    N. Piasco, D. Sidib ´e, C. Demonceaux, and V . Gouet-Brunet, “A survey on visual-based localization: On the benefit of heterogeneous data,” Pattern Recognit., vol. 74, pp. 90–109, 2018

  18. [18]

    Netvlad: CNN architecture for weakly supervised place recognition,

    R. Arandjelovic, P. Gron ´at, A. Torii, T. Pajdla, and J. Sivic, “Netvlad: CNN architecture for weakly supervised place recognition,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 40, no. 6, pp. 1437–1451, 2018

  19. [19]

    Revisit anything: Visual place recognition via image segment re- trieval,

    K. Garg, S. S. Puligilla, S. Kolathaya, K. M. Krishna, and S. Garg, “Revisit anything: Visual place recognition via image segment re- trieval,” inECCV (68), ser. Lecture Notes in Computer Science, vol. 15126, 2024, pp. 326–343

  20. [20]

    Anyloc: Towards universal visual place recognition,

    N. V . Keetha, A. Mishra, J. Karhade, K. M. Jatavallabhula, S. A. Scherer, K. M. Krishna, and S. Garg, “Anyloc: Towards universal visual place recognition,”IEEE Robotics Autom. Lett., vol. 9, no. 2, pp. 1286–1293, 2024

  21. [21]

    Improving image-based localiza- tion by active correspondence search,

    T. Sattler, B. Leibe, and L. Kobbelt, “Improving image-based localiza- tion by active correspondence search,” inECCV (1), vol. 7572, 2012, pp. 752–765

  22. [22]

    Openstreetmap: Challenges and opportunities in machine learning and remote sensing,

    J. E. Vargas-Munoz, S. Srivastava, D. Tuia, and A. X. Falcao, “Openstreetmap: Challenges and opportunities in machine learning and remote sensing,”IEEE Geoscience and Remote Sensing Magazine, vol. 9, no. 1, pp. 184–199, 2020

  23. [23]

    Openstreetslam: Global vehicle localization using openstreetmaps,

    G. Floros, B. van der Zander, and B. Leibe, “Openstreetslam: Global vehicle localization using openstreetmaps,” inICRA, 2013, pp. 1054– 1059

  24. [24]

    Localization in urban environments by matching sensor data to map information,

    C. Mandel and O. Birbach, “Localization in urban environments by matching sensor data to map information,” inEuropean Conference on Mobile Robots, 2013, pp. 50–55

  25. [25]

    Global outer-urban navigation with open- streetmap,

    B. Suger and W. Burgard, “Global outer-urban navigation with open- streetmap,” inICRA, 2017, pp. 1417–1422

  26. [26]

    Do you see the bakery? leveraging geo-referenced texts for global localization in public maps,

    N. Radwan, G. D. Tipaldi, L. Spinello, and W. Burgard, “Do you see the bakery? leveraging geo-referenced texts for global localization in public maps,” inICRA, 2016, pp. 4837–4842

  27. [27]

    Visual cross-view metric localization with dense uncertainty estimates,

    Z. Xia, O. Booij, M. Manfredi, and J. F. Kooij, “Visual cross-view metric localization with dense uncertainty estimates,” inECCV, 2022, pp. 90–106

  28. [28]

    Leveraging deep visual descriptors for hierarchical efficient localization,

    P. Sarlin, F. Debraine, M. Dymczyk, and R. Siegwart, “Leveraging deep visual descriptors for hierarchical efficient localization,” inCoRL, vol. 87, 2018, pp. 456–465

  29. [29]

    From coarse to fine: Robust hierarchical localization at large scale,

    P. Sarlin, C. Cadena, R. Siegwart, and M. Dymczyk, “From coarse to fine: Robust hierarchical localization at large scale,” inCVPR, 2019, pp. 12 716–12 725

  30. [30]

    Maplocnet: Coarse-to-fine feature registration for visual re- localization in navigation maps,

    H. Wu, Z. Zhang, S. Lin, X. Mu, Q. Zhao, M. Yang, and T. Qin, “Maplocnet: Coarse-to-fine feature registration for visual re- localization in navigation maps,” inIROS, 2024, pp. 13 198–13 205

  31. [31]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly,et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” inICLR, 2021

  32. [32]

    Emerging properties in self-supervised vision trans- formers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision trans- formers,” inICCV, 2021, pp. 9630–9640

  33. [33]

    Learning transferable visual models from natural language supervi- sion,

    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 supervi- sion,” inICML, vol. 139, 2021, pp. 8748–8763

  34. [34]

    Splicing vit features for semantic appearance transfer,

    N. Tumanyan, O. Bar-Tal, S. Bagon, and T. Dekel, “Splicing vit features for semantic appearance transfer,” inCVPR, 2022, pp. 10 738– 10 747

  35. [35]

    Styleclip: Text-driven manipulation of stylegan imagery,

    O. Patashnik, Z. Wu, E. Shechtman, D. Cohen-Or, and D. Lischinski, “Styleclip: Text-driven manipulation of stylegan imagery,” inICCV, 2021, pp. 2065–2074

  36. [36]

    Extract free dense labels from CLIP,

    C. Zhou, C. C. Loy, and B. Dai, “Extract free dense labels from CLIP,” inECCV, vol. 13688, 2022, pp. 696–712

  37. [37]

    Clip-dinoiser: Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation,

    M. Wysoczanska, O. Sim ´eoni, M. Ramamonjisoa, A. Bursuc, T. Trzcinski, and P. P´erez, “Clip-dinoiser: Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation,” inECCV, vol. 15119, 2024, pp. 320–337

  38. [38]

    Deep vit features as dense visual descriptors,

    S. Amir, Y . Gandelsman, S. Bagon, and T. Dekel, “Deep vit features as dense visual descriptors,”arXiv preprint arXiv:2112.05814, 2021

  39. [39]

    Roma: Robust dense feature matching,

    J. Edstedt, Q. Sun, G. B ¨okman, M. Wadenb ¨ack, and M. Felsberg, “Roma: Robust dense feature matching,” inCVPR, 2024, pp. 19 790– 19 800

  40. [40]

    Omniglue: Generalizable feature matching with foundation model guidance,

    H. Jiang, A. Karpur, B. Cao, Q. Huang, and A. Ara ´ujo, “Omniglue: Generalizable feature matching with foundation model guidance,” in CVPR, 2024, pp. 19 865–19 875

  41. [41]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” inICLR, 2015

  42. [42]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inCVPR, 2016, pp. 770–778

  43. [43]

    Vision transform- ers need registers,

    T. Darcet, M. Oquab, J. Mairal, and P. Bojanowski, “Vision transform- ers need registers,” inICLR, 2024

  44. [44]

    Vision meets robotics: The KITTI dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The KITTI dataset,”Int. J. Robotics Res., vol. 32, no. 11, pp. 1231– 1237, 2013

  45. [45]

    Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image,

    Y . Shi and H. Li, “Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image,” inCVPR, 2022, pp. 16 989– 16 999

  46. [46]

    You are here: Geolocation by embedding maps and images,

    N. Samano, M. Zhou, and A. Calway, “You are here: Geolocation by embedding maps and images,” inECCV, vol. 12368, 2020, pp. 502–518

  47. [47]

    Back to the feature: Learning robust camera localization from pixels to pose,

    P. Sarlin, A. Unagar, M. Larsson, H. Germain, C. Toft, V . Larsson, M. Pollefeys, V . Lepetit, L. Hammarstrand, F. Kahl, and T. Sattler, “Back to the feature: Learning robust camera localization from pixels to pose,” inCVPR, 2021, pp. 3247–3257

  48. [48]

    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,” inCVPR, 2020, pp. 4063–4071

  49. [49]

    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,” inCVPR, 2021, pp. 3640– 3649