Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

OSMLoc: Single Image-Based Visual Localization in OpenStreetMap with Fused Geometric and Semantic Guidance

T0 review · 4 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read OSMLoc fuses geometry and semantics to localize a single photo against OpenStreetMap, beating prior methods on unseen cities.

desk verdict Solid engineering with a genuinely useful new cross-city benchmark; the depth supervision story needs direct evidence before the geometric contribution is credible. read the letter →

arxiv 2411.08665 v3 pith:2EOCC4NZ submitted 2024-11-13 cs.CV

Youqi Liao , Xieyuanli Chen , Shuhao Kang , Jianping Li , Zhen Dong , Hongchao Fan , Bisheng Yang This is my paper · ORCID
classification cs.CV
keywords visuallocalizationOpenStreetMapcross-modalmatchingbird's-eyeviewdepthdistributionadapterfoundationmodelsemanticalignmentcross-areageneralization
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 proposes OSMLoc, a method that takes one photo and a noisy position guess and returns a three-degree-of-freedom pose (x, y, heading) by matching the photo against a rasterized OpenStreetMap tile. The authors argue that bridging the gap between ground-level imagery and OSM's compact vector map requires fusing two cues the human navigation system uses: geometry and semantics. They therefore combine a frozen foundation-model backbone (DINOv2 from Depth Anything) with a depth distribution adapter that supervises the camera-to-BEV lifting, and an auxiliary semantic alignment loss that pushes the BEV image features toward OSM semantic embeddings. Experiments on the MGL dataset, a newly collected cross-area and cross-condition benchmark covering Detroit, Munich, Taipei, and Brisbane, and the KITTI dataset show OSMLoc outperforming existing single-image image-to-OSM methods, with OSMLoc-B reaching 59.40% PR@5m on MGL versus 54.01% for OrienterNet, and 63.01% versus 46.24% on Detroit without fine-tuning. This matters because OSM is free and global, so a method that localizes against it could reduce dependence on expensive geo-referenced imagery and point-cloud maps.

What carries the argument

The depth distribution adapter (DDA) is the central mechanism. It takes the image feature for each pixel, predicts a distribution over D depth samples along the camera ray, and scatters the feature into a polar BEV grid; during training it is supervised by an L1 loss against the normalized disparity from Depth Anything's decoder. A companion semantic alignment loss aligns the BEV feature at each grid cell with the OSM semantic embedding at the corresponding ground-truth pose location, and the final pose is read out from a softmax-normalized matching score volume over discretized positions and headings.

What would settle it

Take the trained depth distribution adapter and compare its predicted metric depth against LiDAR ground truth on KITTI or another depth-annotated sequence, then deliberately add a constant shift or scale to all depth distributions during inference; if localization recall does not change under that perturbation, the geometric guidance is not providing the metric structure the paper claims, and if a few meters of depth error collapses PR@5m, the claimed generalization depends on metric depth accuracy.

Watch

Extended reading notes

Core claim

The central claim is that fusing geometric and semantic guidance makes single-image localization against OpenStreetMap accurate enough to generalize to cities and camera conditions never seen in training. The paper shows this by replacing the ResNet encoder of OrienterNet with the frozen DINOv2 encoder of Depth Anything, converting the pseudo-ground-truth normalized disparity from Depth Anything into supervision for a per-pixel depth distribution, and scattering the image features into a polar BEV feature map whose cells are then aligned with OSM semantic embeddings at the ground-truth pose. The resulting pose probability volume is produced by dense inner-product matching between the BEV feature and the map feature. The reported numbers are the evidence for the claim: OSMLoc-B beats OrienterNet on position and orientation recall at every threshold on MGL, on all four cities of the CC benchmark, and on KITTI without fine-tuning, and it also improves sequential localization when plugged into a Monte Carlo particle filter as the observation model.

Load-bearing premise

The load-bearing premise is that the method's depth teaching, which only tells the network which objects are nearer or farther and never gives real distances in meters, is enough for the network to build a bird's-eye view laid out in real half-meter intervals; if the network cannot recover true metric depth from that ordinal clue, the whole scene alignment is built on distorted geometry.

Editorial extensions

If this is right

  • If the central claim holds, a single image plus a rough GPS guess is enough to localize against freely available OSM data in unseen cities, removing the need for task-specific fine-tuning on the target area.
  • The same framework can serve as the observation model in a particle filter, and the paper reports that sequential localization further lifts PR@5m on MGL from 59.40% to 77.42%.
  • The KITTI results, obtained without training on KITTI, suggest that cars and robots can use the method as a self-driving localization backend even when no HD map or dense point cloud of the area is available.
  • Because the image encoder is frozen and only lightweight adapters are trained, the method inherits the generalization properties of the foundation model, which the paper identifies as the reason the accuracy gains do not come from a larger architecture.
  • The new cross-area and cross-condition benchmark provides a reusable testbed for measuring whether future I2O localization methods generalize rather than memorize.

Reading between the lines

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

  • Beyond the paper: an unstated consequence is that because the supervision signal is only normalized disparity, the network itself must supply metric scale for the 0.5 m BEV bins; if it silently learns a scene-specific scale rather than a universal one, the reported generalization gains may be tied to the distribution of the training cities rather than to genuine geometric understanding.
  • Beyond the paper: a direct extension would be to swap the normalized-disparity prior for a metric depth prior (or LiDAR during training) and compare recall; a metric supervision version should improve or match OSMLoc's results if metric geometry is what carries the gains.
  • Beyond the paper: the same semantic-alignment trick could transfer to image-to-aerial localization, where satellite imagery or aerial semantic maps provide a comparable cross-view semantic constraint; the paper does not test this setting.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. OSMLoc proposes a single-image visual localization method against OpenStreetMap (OSM) raster tiles. The method extracts image features with a frozen DINOv2 encoder from Depth Anything, lifts them to a polar/cartesian BEV representation using a learned depth distribution supervised by normalized disparity from Depth Anything, and matches the BEV feature against a learned map feature via dense neural matching. A semantic alignment loss aligns the BEV feature to OSM semantic embeddings at the ground-truth pose during training. The authors also introduce a cross-area/cross-condition (CC) benchmark with 7,164 images from Detroit, Munich, Taipei, and Brisbane. Experiments compare OSMLoc-S/B against OrienterNet and several I2A/I2I baselines on the MGL dataset, the new CC benchmark, and KITTI, reporting consistently higher position/orientation recall, plus sequential localization results with a particle filter.

Significance. If the claims hold, the paper makes a useful contribution to image-to-OSM localization: it is one of the first to show that a visual foundation model (DINOv2/Depth Anything) can be combined with a geometric BEV lifting and semantic alignment to outperform the prior state of the art (OrienterNet) across same-area and cross-area benchmarks. The open-sourced code, pre-trained models, and new CC benchmark are tangible assets for the community, and the paper includes honest failure-mode analysis and a sequential-localization extension. The main significance is conditional on the geometry-guided contribution being actually load-bearing; the reported ablations for geometric guidance alone are small and inconsistent, and the paper does not directly validate metric depth accuracy, which is the central mechanism of that contribution.

major comments (4)
  1. [§3.2 and §5.1] The normalized-disparity supervision in Eq. (4)–(5) is invariant to per-image affine rescaling of inverse depth, yet the BEV lifting scatters features into metric bins with d0=0 and Δ=0.5 m. The network must therefore recover absolute metric depth, but the only supervision that can provide metric scale is indirect (pose and semantic losses applied after the BEV projection). The paper reports no direct evaluation of metric depth accuracy. This is load-bearing for contribution #2; please add a quantitative evaluation of the predicted depth distribution against metric ground truth (e.g., on KITTI) or otherwise show that the normalized-disparity prior determines metric BEV geometry.
  2. [Table 6] The ablation of geometric guidance alone (variant [B] vs [A]) shows only a small and inconsistent effect: PR@1m improves from 11.52 to 12.33 but PR@5m decreases from 55.53 to 55.15, while semantic guidance alone ([C]) gives a larger PR@1m gain. This weakens the claim that geometric guidance is a primary source of the reported gains. Please discuss this pattern and, ideally, provide an ablation where the geometric prior is replaced with a metric depth prior (e.g., Depth Pro or LiDAR-derived depth) to isolate the role of metric scale.
  3. [Tables 2–4 and 8–9] No error bars, confidence intervals, or significance tests are reported. The central comparisons (e.g., OSMLoc-B vs OrienterNet on the MGL and CC benchmarks) are based on single training runs, so the claimed statistical superiority is not established. Given the scale of the reported differences this is not obviously fatal, but the paper should report at least 2–3 seeds with mean and standard deviation, or a paired test, for the main tables.
  4. [§5.1, KITTI protocol] The KITTI evaluation restricts inference to pose candidates within ±20 m and ±10° of the ground-truth pose, which is the OrienterNet protocol. The paper's language in the abstract and introduction ('generalization capability') could be read as implying global localization; please explicitly state that the KITTI result is a local-refinement task with an informative prior, and clarify how this differs from the global pose search performed on MGL/CC.
minor comments (4)
  1. [§3.3] Equation numbering is inconsistent: Eq. (6) is used for the semantic alignment loss, the matching score, the pose loss, the joint loss, and the inference argmax. Please renumber.
  2. [§3.3, semantic alignment] The semantic alignment loss is described as 'self-supervision,' but it requires the ground-truth pose. This is more accurately a pose-supervised auxiliary loss; please adjust the wording.
  3. [§5.2, CC benchmark] The statement that Taipei and Brisbane have 'limited useful information' is speculative; consider quantifying this, e.g., with per-city feature statistics or failure localization analysis.
  4. [§2.2] The related-work discussion of MapLocNet mentions HD-map dependence but the paper does not compare against MapLocNet in the experiments. If code or numbers are unavailable, it would be helpful to state that explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: OSMLoc's gains are empirical and rest on external depth/semantic priors, with no fitted parameter renamed as prediction.

full rationale

I found no circular step in OSMLoc. The geometric guidance is distilled from an external foundation model, Depth Anything, via the normalized-disparity loss in Eqs. (4)-(5); the semantic targets are rasterized OpenStreetMap data; and the CC benchmark and KITTI evaluations use scenes not in the training set. The reader-identified concern that Eq. (4) removes metric scale while the BEV transform uses 0.5 m bins is a verification/calibration weakness, not a circularity: the network still predicts a metric depth distribution and is supervised by an external ordinal prior, so the output is not defined as its input and no fitted parameter is renamed as a prediction. The semantic alignment loss of Eq. (6) uses the ground-truth pose during training, but that is standard supervised training, not an entanglement of the evaluation. The paper's improvements are reported empirically through ablations and comparisons against independent baselines, and the few self-citations (e.g., prior sequential-localization work) are background references, not load-bearing justifications. The derivation chain is therefore self-contained with respect to the external benchmarks and priors.

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

The paper introduces no new physical or theoretical entities. Its load-bearing assumptions are the gravity alignment, the transfer of ordinal depth priors to metric BEV lifting, and the sufficiency of OSM rasterization. The main free parameters are loss weights and geometry discretization choices selected for the experiments.

free parameters (4)
  • Loss weights (lambda1, lambda2, lambda3) = (1, 20, 10)
    Selected through ablations in Table 8 on the MGL validation set; directly controls the balance of pose, disparity, and semantic losses.
  • BEV depth grid (D, L, d0, Delta) = D=64, L=129, d0=0, Delta=0.5 m
    Manual quantization of the camera frustum into metric distance bins; sets the spatial resolution and maximum range of BEV lifting.
  • Orientation sampling K = 64 train, 256 inference
    Discretization of yaw space for the pose volume; larger K improves orientation resolution at inference.
  • Particle filter settings (Np, sigma) = Np=1000/200, sigma=2.0
    Sequential localization MCL hyperparameters chosen for all experiments; affects smoothing but not the single-image claim.
assumptions (5)
  • domain assumption The camera is gravity-aligned and roll/pitch are rectified to zero using an IMU or gyroscope.
    Sec. 3 states this assumption; the BEV polar strip model and 3-DoF pose parameterization rely on it.
  • domain assumption Normalized disparity from Depth Anything is a valid pseudo-ground-truth for supervising the depth distribution.
    Sec. 3.2, Eq. (4)-(5); the method distills ordinal depth while the BEV transform needs metric depth, and no evidence is given that ordinal supervision yields metric accuracy.
  • domain assumption OSM rasterization at 50 cm/pixel with the listed areas/ways/nodes categories preserves enough semantic structure for localization.
    Sec. 3 and Fig. 2; the map encoder and semantic alignment only see this rasterization.
  • domain assumption Mapillary ground-truth poses for MGL and the CC benchmark are accurate enough to serve as labels.
    Sec. 5.1 describes MGL poses from fusion of motion pattern and GPS; the CC benchmark inherits this pipeline, and no pose-quality analysis is included.
  • standard math Standard deep learning toolkit assumptions: SGD, BCE pose loss, softmax over pose volume.
    No unusual math is used; losses are standard and not load-bearing beyond training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OSMLoc: Single Image-Based Visual Localization in OpenStreetMap with Fused Geometric and Semantic Guidance." pith.science (2026). https://pith.science/paper/2EOCC4NZ

@misc{pith2026241108665,
  author       = {Pith},
  title        = {Pith review of: OSMLoc: Single Image-Based Visual Localization in OpenStreetMap with Fused Geometric and Semantic Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2EOCC4NZ}},
  note         = {Machine review of arXiv:2411.08665}
}
read the original abstract

OpenStreetMap (OSM), a rich and versatile source of volunteered geographic information (VGI), facilitates human self-localization and scene understanding by integrating nearby visual observations with vectorized map data. However, the disparity in modalities and perspectives poses a major challenge for effectively matching camera imagery with compact map representations, thereby limiting the full potential of VGI data in real-world localization applications. Inspired by the fact that the human brain relies on the fusion of geometric and semantic understanding for spatial localization tasks, we propose the OSMLoc in this paper. OSMLoc is a brain-inspired visual localization approach based on first-person-view images against the OSM maps. It integrates semantic and geometric guidance to significantly improve accuracy, robustness, and generalization capability. First, we equip the OSMLoc with the visual foundational model to extract powerful image features. Second, a geometry-guided depth distribution adapter is proposed to bridge the monocular depth estimation and camera-to-BEV transform. Thirdly, the semantic embeddings from the OSM data are utilized as auxiliary guidance for image-to-OSM feature matching. To validate the proposed OSMLoc, we collect a worldwide cross-area and cross-condition (CC) benchmark for extensive evaluation. Experiments on the MGL dataset, CC validation benchmark, and KITTI dataset have demonstrated the superiority of our method. Code, pre-trained models, CC validation benchmark, and additional results are available at: https://github.com/WHU-USI3DV/OSMLoc.

Figures

Figures reproduced from arXiv: 2411.08665 by the authors.

Figure 1
Figure 1. Image-to-OpenstreetMap localization aims to es￾timate the 3 degrees of freedom (3-DoF) camera pose. (a) shows the core idea of our method, which fuses geometric and semantic guidance into the framework. (b) shows the global evaluation results. can utilize geometric cues from their surroundings for orien￾tation and localization. Building on this, [2, 23] showed that the retrosplenial–hippocampal circuit in the brain … view at source ↗
Figure 3
Figure 3. Workflow of the OSMLoc. (Sec 3.3) module estimates pose by dense feature matching. The semantic alignment is employed as an auxiliary task to improve the scene understanding ability of the image model. Pose loss 𝑝𝑜𝑠𝑒, disparity loss 𝑑𝑖𝑠 and semantic alignment loss 𝑠𝑒𝑚 are leveraged to supervise the network jointly. 3.1. Feature extraction with foundation model For the image model, we adopt the DINOv2 [25] encoder… view at source ↗
Figure 2
Figure 2. Illustration of the OSM pre-processing. Since our approach incorporates cross-view and cross￾modality data, the image and OSM data require pre-processing before being fed into the framework. For the image , we rectify the roll and pitch angles to zero using the known gravity direction and ensure the principal axis is horizontal. The OSM data is stored in structured metadata and repre￾sents the areas, ways and nodes… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Illustration of the depth distribution adapter. For each pixel, we predict the depth distribution 𝜶 ∈ ℝ𝐷 and scatter the image feature 𝑭𝑖𝑚𝑔 to the BEV feature 𝑭𝑏𝑒𝑣 by the corresponding depth distribution 𝜶. During training, the depth distribution prediction is transfer…
Figure 5
Figure 5. Figure 5: OSM data coverage and image trajectories in the CC validation benchmark. The red points represent the locations of the images [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Query image and map tile pairs from the CC validation benchmark. Red arrow →represents the GT pose in the map tile. • MGL dataset: consists of 2580 sequences with ap￾proximately 828 K images collected in 13 cities from the Mapillary platform, which exposes the camera c…
Figure 7
Figure 7. Figure 7: Position and orientation recall curves of the top candidates on the MGL dataset. The gray, orange, and navy curves are for OrienterNet, OSMLoc-S and OSMLoc-B respectively. the test areas of Taipei and Brisbane are situated in central urban districts with numerous skysc…
Figure 8
Figure 8. Figure 8: Qualitative results of OSMLoc-S on the CC validation benchmark. (a) Detroit. (b) Munich. (c) Taipei. (d) Brisbane. Black arrow →represents the predicted pose and red arrow →represents the GT pose in the map [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Qualitative results on the MGL dataset, CC validation benchmark and KITTI dataset. The red rectangle represents the wrong localization result and the green rectangle represents the correct localization result. Black arrow →represents the predicted pose and red arrow →r…
Figure 10
Figure 10. Figure 10: Failure cases. The first and second ones are from the MGL dataset and the CC validation benchmark, while the third and last ones are from the KITTI dataset. Black arrow →represents the predicted pose and red arrow →represents the GT pose in the map. under different co…
Figure 11
Figure 11. Figure 11: Qualitative results of sequential localization in Munich. (a) comparable localization results of single-image and sequential images. (b) Incorporating multi-frame observation and the motion pattern reduces the ambiguity and consistently improves the accuracy and robus…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Aerial-ground Cross-modal Localization: Dataset, Ground-truth, and Benchmark

    cs.RO 2025-09 conditional novelty 6.0 of 10

    A large-scale dataset and benchmark for localizing ground images against aerial ALS point clouds in three cities, with ground-truth poses generated by MLS-ALS registration.

  2. OPAL: Visibility-aware LiDAR-to-OpenStreetMap Place Recognition via Adaptive Radial Fusion

    cs.CV 2025-04 conditional novelty 6.0 of 10

    OPAL matches single LiDAR scans to OpenStreetMap tiles using deterministic visibility masks and learnable radial attention, improving top-1 recall by up to 15.98% on KITTI at real-time speed.

Reference graph

Works this paper leans on

72 extracted references · 67 canonical work pages · cited by 2 Pith papers

  1. [1]

    L.Hermer,E.S.Spelke, Ageometricprocessforspatialreorientation in young children, Nature 370 (1994) 57–59

    work page 1994

  2. [2]

    R. A. Epstein, L. K. Vass, Neural systems for landmark-based wayfinding in humans, Philosophical Transactions of the Royal Society B: Biological Sciences 369 (2014) 20120533

    work page 2014

  3. [3]

    J. Yuan, J. Zhang, S. Ding, X. Dong, Cooperative localization for disconnected sensor networks and a mobile robot in friendly environments, Information Fusion 37 (2017) 22–36

    work page 2017

  4. [4]

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

    work page 2021

  5. [5]

    H. Yang, J. Yuan, Y. Gao, X. Sun, X. Zhang, Uplp-slam: Unified point-line-plane feature fusion for rgb-d visual slam, Information Fusion 96 (2023) 51–65

    work page 2023

  6. [6]

    K.Zhou,J.Wang,W.Xu,L.Song,Z.Ye,C.Guo,C.Li, Learningand grounding visual multimodal adaptive graph for visual navigation, Information Fusion (2025) 103009

    work page 2025

  7. [7]

    5297–5307

    R.Arandjelovic,P.Gronat,A.Torii,T.Pajdla,J.Sivic, Netvlad:Cnn architectureforweaklysupervisedplacerecognition, in:Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 5297–5307

    work page 2016

  8. [8]

    Piras, G

    L. Piras, G. Giacinto, Information fusion in content based image retrieval: A comprehensive overview, Information Fusion 37 (2017) 50–60

    work page 2017

Show all 72 references
  1. [9]

    Hausler, S

    S. Hausler, S. Garg, M. Xu, M. Milford, T. Fischer, Patch-netvlad: Multi-scalefusionoflocally-globaldescriptorsforplacerecognition, in:ProceedingsoftheIEEE/CVFconferenceoncomputervisionand pattern recognition, 2021, pp. 14141–14152

    2021

  2. [10]

    Y.Li,J.Ma,Y.Zhang, Imageretrievalfromremotesensingbigdata: A survey, Information Fusion 67 (2021) 94–115

    2021

  3. [11]

    S. Yu, C. Wang, Z. Yu, X. Li, M. Cheng, Y. Zang, Deep regression for lidar-based localization in dense urban areas, ISPRS Journal of Photogrammetry and Remote Sensing 172 (2021) 240–252

    2021

  4. [12]

    K. Chen, H. Yu, W. Yang, L. Yu, S. Scherer, G.-S. Xia, I2d-loc: Camera localization via image to lidar depth flow, ISPRS Journal of Photogrammetry and Remote Sensing 194 (2022) 209–221

    2022

  5. [13]

    C. Shi, J. Li, J. Gong, B. Yang, G. Zhang, An improved lightweight deep neural network with knowledge distillation for local feature extractionandvisuallocalizationusingimagesandlidarpointclouds, ISPRS journal of photogrammetry and remote sensing 184 (2022) 177–188

    2022

  6. [14]

    X.Zou,J.Li,Y.Wang,F.Liang,W.Wu,H.Wang,B.Yang,Z.Dong, Patchaugnet: Patch feature augmentation-based heterogeneous point cloudplacerecognitioninlarge-scalestreetscenes, ISPRSJournalof Photogrammetry and Remote Sensing 206 (2023) 273–292

    2023

  7. [15]

    Y. Wang, W. Jiao, H. Fan, G. Zhou, A framework for fully auto- matedreconstructionofsemanticbuildingmodelaturban-scaleusing textured lod2 data, ISPRS Journal of Photogrammetry and Remote Sensing 216 (2024) 90–108

    2024

  8. [16]

    URL: https://planet

    OSM_Foundation, Osm static, 2024. URL: https://planet. openstreetmap.org/statistics/data_stats.html

    2024

  9. [17]

    H. Fan, A. Zipf, Q. Fu, P. Neis, Quality assessment for building foot- prints data on openstreetmap, International Journal of Geographical Information Science 28 (2014) 700–719

    2014

  10. [18]

    J. E. Vargas-Munoz, S. Srivastava, D. Tuia, A. X. Falcao, Open- streetmap: Challenges and opportunities in machine learning and remote sensing, IEEE Geoscience and Remote Sensing Magazine 9 (2020) 184–199

    2020

  11. [19]

    Samano, M

    N. Samano, M. Zhou, A. Calway, You are here: Geolocation by embeddingmapsandimages,in:ComputerVision–ECCV2020:16th European Conference, Glasgow, UK, August 23–28, 2020, Proceed- ings, Part XXIII 16, Springer, 2020, pp. 502–518

    2020

  12. [20]

    M. Zhou, X. Chen, N. Samano, C. Stachniss, A. Calway, Efficient localisation using images and openstreetmaps, in: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE, 2021, pp. 5507–5513

    2021

  13. [21]

    Sarlin, D

    P.-E. Sarlin, D. DeTone, T.-Y. Yang, A. Avetisyan, J. Straub, T. Mal- isiewicz,S.R.Bulo,R.Newcombe,P.Kontschieder,V.Balntas, Ori- enternet:Visuallocalizationin2dpublicmapswithneuralmatching, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,...

    2023

  14. [22]

    H.Wu, Z.Zhang, S.Lin, X.Mu, Q.Zhao, M.Yang,T. Qin, Maploc- net: Coarse-to-fine feature registration for visual re-localization in navigation maps, in: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE, 2024, pp. 13198– 13205

    2024

  15. [23]

    R. A. Epstein, E. Z. Patai, J. B. Julian, H. J. Spiers, The cognitive map in humans: spatial navigation and beyond, Nature neuroscience 20 (2017) 1504–1513

    2017

  16. [24]

    P. Foo, W. H. Warren, A. Duchon, M. J. Tarr, Do humans integrate routes into a cognitive map? map-versus landmark-based navigation of novel shortcuts., Journal of Experimental Psychology: Learning, Memory, and Cognition 31 (2005) 195

    2005

  17. [25]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. HAZIZA, F. Massa, A. El-Nouby, et al., Dinov2: Learning robust visual features without supervision, Transactions on Machine Learning Research (2024)

    2024

  18. [26]

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, H. Zhao, Depth anything: Unleashing the power of large-scale unlabeled data, in: Liao et al.: Preprint submitted to Elsevier Page 15 of 17 OSMLoc Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024...

    2024

  19. [27]

    Sattler, B

    T. Sattler, B. Leibe, L. Kobbelt, Fast image-based localization using direct 2d-to-3d matching, in: 2011 International Conference on Computer Vision, IEEE, 2011, pp. 667–674

    2011

  20. [28]

    Cheng, W

    W. Cheng, W. Lin, X. Zhang, M. Goesele, M.-T. Sun, A data- driven point cloud simplification framework for city-scale image- basedlocalization, IEEETransactionsonImageProcessing26(2016) 262–275

    2016

  21. [29]

    Q. Li, J. Zhu, J. Liu, R. Cao, H. Fu, J. M. Garibaldi, Q. Li, B. Liu, G. Qiu, 3d map-guided single indoor image localization refinement, ISPRS Journal of Photogrammetry and Remote Sensing 161 (2020) 13–26

    2020

  22. [30]

    T.-Y.Lin,S.Belongie,J.Hays, Cross-viewimagegeolocalization, in: ProceedingsoftheIEEEConferenceonComputerVisionandPattern Recognition, 2013, pp. 891–898

    2013

  23. [31]

    Y. Tian, C. Chen, M. Shah, Cross-view image matching for geo- localization in urban environments, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 3608–3616

    2017

  24. [32]

    P.-E.Sarlin,E.Trulls,M.Pollefeys,J.Hosang,S.Lynen, Snap:Self- supervised neural maps for visual positioning and semantic under- standing, Advances in Neural Information Processing Systems 36 (2024)

    2024

  25. [33]

    G. Li, M. Qian, G.-S. Xia, Unleashing unlabeled data: A paradigm for cross-view geo-localization, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16719–16729

    2024

  26. [34]

    Civera, Mapillary street-level sequences: A dataset for lifelong place recognition, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp

    F.Warburg,S.Hauberg,M.Lopez-Antequera,P.Gargallo,Y.Kuang, J. Civera, Mapillary street-level sequences: A dataset for lifelong place recognition, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2626–2635

    2020

  27. [35]

    URL:https://kartaview.org/

    Grab_Holdings, Kartaview, 2024. URL:https://kartaview.org/

    2024

  28. [36]

    S. Unar, X. Wang, C. Zhang, Visual and textual information fusion using kernel method for content based image retrieval, information Fusion 44 (2018) 176–187

    2018

  29. [37]

    K. T. Ahmed, S. Ummesafi, A. Iqbal, Content based image retrieval using image features information fusion, Information Fusion 51 (2019) 76–99

    2019

  30. [38]

    Z. Li, C. D. W. Lee, B. X. L. Tung, Z. Huang, D. Rus, M. H. Ang,Hot-netvlad:Learningdiscriminatorykeypointsforvisualplace recognition, IEEE Robotics and Automation Letters 8 (2023) 974– 980

    2023

  31. [39]

    S. Zhu, L. Yang, C. Chen, M. Shah, X. Shen, H. Wang, R2former: Unified retrieval and reranking transformer for place recognition, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 19370–19380

    2023

  32. [40]

    Dusmanu, I

    M. Dusmanu, I. Rocco, T. Pajdla, M. Pollefeys, J. Sivic, A. Torii, T. Sattler, D2-net: A trainable cnn for joint description and detection of local features, in: Proceedings of the ieee/cvf conference on computer vision and pattern recognition, 2019, pp. 8092–8101

    2019

  33. [41]

    Sarlin, D

    P.-E. Sarlin, D. DeTone, T. Malisiewicz, A. Rabinovich, Superglue: Learning feature matching with graph neural networks, in: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 4938–4947

    2020

  34. [42]

    2938–2946

    A.Kendall,M.Grimes,R.Cipolla, Posenet:Aconvolutionalnetwork forreal-time6-dofcamerarelocalization, in:ProceedingsoftheIEEE international conference on computer vision, 2015, pp. 2938–2946

    2015

  35. [43]

    K. Liu, Q. Li, G. Qiu, Posegan: A pose-to-image translation frame- workforcameralocalization, ISPRSJournalofPhotogrammetryand Remote Sensing 166 (2020) 308–315

    2020

  36. [44]

    S. Tang, Y. Li, J. Wan, Y. Li, B. Zhou, R. Guo, W. Wang, Y. Feng, Transcnnloc: End-to-end pixel-level learning for 2d-to-3d pose esti- mationindynamicindoorscenes, ISPRSJournalofPhotogrammetry and Remote Sensing 207 (2024) 218–230

    2024

  37. [45]

    Floros, B

    G. Floros, B. Van Der Zander, B. Leibe, Openstreetslam: Global ve- hicle localization using openstreetmaps, in: 2013 IEEE international conference on robotics and automation, IEEE, 2013, pp. 1054–1059

    2013

  38. [46]

    C. Dong, Z. Hong, S. Li, L. Hu, H. Gao, Augmenting vision with radar for all-weather geo-localization without a prior hd map, in: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE, 2024, pp. 3305–3311

    2024

  39. [47]

    Z.Zhou,Z.Qi,L.Cheng,G.Xiong, Seglocnet:Multimodallocaliza- tion network for autonomous driving via bird’s-eye-view segmenta- tion, arXiv preprint arXiv:2502.20077 (2025)

    2025 arXiv

  40. [48]

    F. Yan, O. Vysotska, C. Stachniss, Global localization on open- streetmap using 4-bit semantic descriptors, in: 2019 European con- ference on mobile robots (ECMR), IEEE, 2019, pp. 1–7

    2019

  41. [49]

    Y.Cho,G.Kim,S.Lee,J.-H.Ryu, Openstreetmap-basedlidarglobal localization in urban environment without a prior lidar map, IEEE Robotics and Automation Letters 7 (2022) 4999–5006

    2022

  42. [50]

    Lee, J.-H

    S. Lee, J.-H. Ryu, Autonomous vehicle localization without prior high-definition map, IEEE Transactions on Robotics (2024)

    2024

  43. [51]

    S. Kang, M. Y. Liao, Y. Xia, O. Wysocki, B. Jutzi, D. Cremers, Opal: Visibility-aware lidar-to-openstreetmap place recognition via adaptive radial fusion, arXiv preprint arXiv:2504.19258 (2025)

    2025 arXiv

  44. [52]

    Z. Liu, H. Tang, A. Amini, X. Yang, H. Mao, D. L. Rus, S. Han, Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation, in: 2023 IEEE international conference on robotics and automation (ICRA), IEEE, 2023, pp. 2774–2781

    2023

  45. [53]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T.Unterthiner,M.Dehghani,M.Minderer,G.Heigold,S.Gelly,etal., An image is worth 16x16 words: Transformers for image recognition at scale, in: International Conference on Learning Representations, 2021

    2021

  46. [54]

    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 (2023)

    2023

  47. [55]

    Mirjalili, M

    R. Mirjalili, M. Krawez, W. Burgard, Fm-loc: Using foundation models for improved vision-based localization, in: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE, 2023, pp. 1381–1387

    2023

  48. [56]

    J.Philion,S.Fidler, Lift,splat,shoot:Encodingimagesfromarbitrary camera rigs by implicitly unprojecting to 3d, in: Computer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23– 28, 2020, Proceedings, Part XIV 16, Springer, 2020, pp. 194–210

    2020

  49. [57]

    Lentsch, Z

    T. Lentsch, Z. Xia, H. Caesar, J. F. Kooij, Slicematch: Geometry- guided aggregation for cross-view pose estimation, in: Proceedings oftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition, 2023, pp. 17225–17234

    2023

  50. [58]

    W. Yin, C. Zhang, H. Chen, Z. Cai, G. Yu, K. Wang, X. Chen, C. Shen, Metric3d: Towards zero-shot metric 3d prediction from a single image, in: Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 9043–9053

    2023

  51. [59]

    Bochkovskiy, A

    A. Bochkovskiy, A. Delaunoy, H. Germain, M. Santos, Y. Zhou, S. Richter, V. Koltun, Depth pro: Sharp monocular metric depth in less than a second, in: The Thirteenth International Conference on Learning Representations, 2025

    2025

  52. [60]

    I. A. Barsan, S. Wang, A. Pokrovsky, R. Urtasun, Learning to localizeusingalidarintensitymap,in:ConferenceonRobotLearning (CoRL), 2018

    2018

  53. [61]

    URL: https://www.mapillary

    Mapillary, Mapillary website, 2024. URL: https://www.mapillary. com/

    2024

  54. [62]

    URL:https://www.openstreetmap.org/

    OSM, Osm website, 2024. URL:https://www.openstreetmap.org/

    2024

  55. [63]

    Geiger, P

    A. Geiger, P. Lenz, R. Urtasun, Are we ready for autonomous driving?thekittivisionbenchmarksuite,in:2012IEEEconferenceon computervisionandpatternrecognition,IEEE,2012,pp.3354–3361

    2012

  56. [64]

    Y. Shi, H. Li, Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image, in: Proceedings of the IEEE/CVFConferenceonComputerVisionandPatternRecognition, 2022, pp. 17010–17020

    2022

  57. [65]

    Z. Xia, O. Booij, M. Manfredi, J. F. Kooij, Visual cross-view metric localization with dense uncertainty estimates, in: European Conference on Computer Vision, Springer, 2022, pp. 90–106

    2022

  58. [66]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer Liao et al.: Preprint submitted to Elsevier Page 16 of 17 OSMLoc vision and pattern recognition, 2016, pp. 770–778

    2016

  59. [67]

    Simonyan, A

    K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition, arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  60. [68]

    A. v. d. Oord, Y. Li, O. Vinyals, Representation learning with con- trastive predictive coding, arXiv preprint arXiv:1807.03748 (2018)

    2018 arXiv

  61. [69]

    Zhang, D

    C. Zhang, D. Han, Y. Qiao, J. U. Kim, S.-H. Bae, S. Lee, C. S. Hong, Faster segment anything: Towards lightweight sam for mobile applications, arXiv preprint arXiv:2306.14289 (2023)

    2023 arXiv

  62. [70]

    Hinton, O

    G. Hinton, O. Vinyals, J. Dean, Distilling the knowledge in a neural network, arXiv preprint arXiv:1503.02531 (2015)

    2015 arXiv

  63. [71]

    X. Chen, I. Vizzo, T. Läbe, J. Behley, C. Stachniss, Range image- based lidar localization for autonomous vehicles, in: 2021 IEEE InternationalConferenceonRoboticsandAutomation(ICRA),IEEE, 2021, pp. 5802–5808

    2021

  64. [72]

    Thrun, Probabilistic robotics, Communications of the ACM 45 (2002) 52–57

    S. Thrun, Probabilistic robotics, Communications of the ACM 45 (2002) 52–57. Liao et al.: Preprint submitted to Elsevier Page 17 of 17

    2002

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.