Pith. sign in

REVIEW 2 major objections 4 minor 34 references

A drone can localize itself with meter-level accuracy using only public satellite imagery, bare-earth elevation maps, and building-footprint annotations, even in areas it has never flown before.

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 →

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

T0 review · deepseek-v4-flash

2026-08-02 03:01 UTC pith:HECXXRP2

load-bearing objection Useful single-frame localization contribution, but the headline 55-km GNSS-denied trajectory claim rests on a motion prior derived from the same GNSS reference used for scoring. the 2 major comments →

arxiv 2607.14009 v1 pith:HECXXRP2 submitted 2026-07-15 cs.RO

AeroMap3D: Anchoring Monocular UAV 6-DoF Localization to Visual-Geometric-Semantic Map Priors

Zhiyun Deng , Luis Sentis This is my paper
classification cs.RO
keywords UAV localizationGNSS-denied navigationmap-based pose estimationcross-view image registrationdigital elevation modelbuilding footprint semanticsscale-yaw adaptationdelayed-state EKF
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.

This paper tries to show that a drone flying without GPS can keep itself metrically anchored to the world using only widely available public map layers—satellite imagery, bare-earth elevation models, and building-footprint annotations—rather than site-specific 3D reconstructions or surveyed models. The central problem is that a drone's downward camera sees buildings and roofs while the elevation map gives only ground height, so visually correct matches become geometrically wrong 3D points. The paper's answer is to predict and correct the scale and heading mismatch between drone image and map tile with a lightweight network trained only on synthetic pairs, and to discard matches that fall on known building footprints before lifting the remaining matches onto the terrain model and solving pose with RANSAC-PnP. If true, this means a drone can localize in unexplored urban areas using maps that already exist for large parts of the world, without per-site preparation. The headline result—5.88 m mean 3D error over 55 km—depends on a relative-motion prior derived from the same GNSS reference used for scoring, so the 'GNSS-denied' label should be read as 'map-anchored with a reference-quality motion prior.'

Core claim

The paper's central claim is that two classic failures of map-referenced aerial localization—cross-view scale/yaw mismatch and the bare-earth elevation model's blindness to building height—can be overcome without retraining a dense matcher or building a site-specific 3D model. A lightweight Siamese network trained only on synthetic satellite-derived pairs estimates the scale ratio and yaw offset between the UAV image and a north-up satellite tile, rectifying the pair so a frozen, internet-pretrained dense matcher can register it. Then, before 2D-to-3D lifting, building-footprint annotations are used as a mask to reject correspondences that would inherit a building-height bias from the elevat

What carries the argument

The two load-bearing mechanisms are (1) a lightweight scale–yaw adapter that predicts the altitude- and field-of-view-induced scale ratio and the heading-induced yaw offset between a UAV image and a north-up satellite tile, converting a challenging cross-view pair into a geometry where a frozen dense matcher can find correspondences; and (2) building-footprint masking applied before elevation-model lifting, which removes matches on roofs and facades that would otherwise inject a coherent height bias into RANSAC-PnP. The adapter is what makes the system generalize without retraining; the mask is what makes a bare-earth elevation model a valid 3D reference in urban scenes.

Load-bearing premise

The headline trajectory accuracy relies on a relative-motion prior generated by perturbing the same GNSS reference poses that later define the error metric, so the 5.88 m result is not a GNSS-denied motion estimate—it assumes an oracle-quality odometry prior between map updates.

What would settle it

Run the same pipeline with a real onboard IMU/visual odometry prior (or perturb the relative motion with noise independent of the reference poses) and check whether the 50 m success and 5.88 m mean error persist; if trajectory error degrades to retrieval-level (around 17 m) or worse, the map updates are not the primary anchor. Separately, remove the building-footprint mask on a test site and measure whether single-frame success falls back near the 88.24% dense-PnP baseline.

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

If this is right

  • If the 5.88 m trajectory figure holds, map-anchored monocular localization from public layers is a viable drift-bounding strategy for urban UAVs, with no site survey or prior 3D reconstruction.
  • A frozen dense matcher can be repurposed to a new domain by normalizing just two nuisance parameters, avoiding expensive fine-tuning on target data.
  • Public building-footprint annotations can convert a bare-earth elevation model into a usable 3D reference for urban flight, not just open terrain.
  • The delayed-state EKF formulation means the pipeline can run online, absorbing asynchronous, delayed map-based pose updates without reprocessing the whole trajectory.

Where Pith is reading between the lines

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

  • Because the motion prior in the experiments is derived from the same GNSS reference poses that define the error metric, the 5.88 m number overstates what a true GNSS-denied system would achieve with onboard odometry; the paper's own visual-odometry baseline suggests map updates are doing the anchoring, but the gap between a reference-derived prior and a real IMU/visual prior is untested.
  • The adapter's synthetic-only training suggests a zero-shot transfer recipe for other cross-view registration tasks—ground robots versus satellite, or oblique aerial views—where the same scale/yaw nuisance structure dominates.
  • In cities with sparse or outdated building footprints, the semantic mask will miss structures and the method should degrade toward the 88% single-frame baseline; one testable extension is to generate the mask online with a building detector instead of relying on a pre-existing footprint layer.
  • Because the benchmark provides translation-only ground truth and the authors state the device lacks validated full 6-DoF ground truth, the roll/pitch/yaw components of the claimed 6-DoF poses are the least supported part of the result.

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

2 major / 4 minor

Summary. AeroMap3D proposes a monocular UAV 6-DoF localization system that registers UAV imagery to publicly available satellite imagery, bare-earth DEMs, and OSM footprints. A scale/yaw adapter (trained on synthetic data) normalizes the UAV-map pair before a frozen dense matcher; OSM masks filter out building pixels before DEM lifting and RANSAC-PnP. A delayed-state EKF fuses the resulting map-anchored pose estimates with a relative-motion prior. The paper reports 95.69% single-frame success (up from 88.24% without OSM filtering), 99.2% registration success with the adapter, and a 5.88 m mean trajectory error over 55 km using a reference-derived stochastic motion prior. It also introduces the UAV-Terra3D benchmark.

Significance. If the trajectory result were obtained under a truly GNSS-denied motion prior, the paper would be a significant step toward infrastructure-free UAV localization using only public geospatial data. The single-frame ablation is clean and compelling: disabling only the OSM filter raises success from 88.24% to 95.69% with all other components fixed. The cross-dataset adapter transfer (synthetic training on UAV-VisLoc, testing on UAV-Terra3D) is a notable strength, as is the release of a 55-km benchmark with co-registered visual-geometric-semantic layers. However, the headline trajectory claim is not currently supported as a GNSS-denied result, and the 6-DoF claim is not validated because only translational ground truth is available.

major comments (2)
  1. [Sec. V, Eq. (19); Table III lower half] The reported 'Ours (PnP+EKF)' trajectory result uses a relative-motion prior generated by perturbing the GNSS reference poses (Eq. 19: ΔT̃_t = ΔT*_t Exp(ξ_t)), and the error metric is computed against the same GNSS reference. This is not representative of GNSS-denied flight: the prior carries nearly all the absolute trajectory information, and the map updates only need to correct small injected perturbations. The paper itself acknowledges this as a limitation in Sec. VII, but the abstract and Q4 conclusions present the 5.88 m mean error as evidence for GNSS-denied navigation. The visual-only ORB prior mentioned in Sec. V is not used for the reported AeroMap3D trajectory, so no result with a truly GNSS-independent motion prior is provided. Please supply a trajectory evaluation with the ORB prior (or another real odometry prior) and re-frame the GNSS-denied claim accordingly.
  2. [Sec. VI-D, Table III footnote] The paper claims 6-DoF localization in the title and abstract, but the evaluation is restricted to the 3-DoF position component, as stated in the Table III footnote: 'UA V-Terra3D provides ground truth only for the 3-DoF position component. Therefore, the evaluation is restricted to translational localization accuracy, while rotational accuracy is not reported.' The dataset section also states that the collection device does not provide independently validated full 6-DoF ground truth. Thus, the 6-DoF claim is not supported by the experiments. Either provide orientation error results (even approximate) or limit the claim to 3-DoF position accuracy.
minor comments (4)
  1. [Sec. VI-C, Fig. 5] The registration 'success rate' in Fig. 5 is a pseudo success rate based on geometric verification (valid fundamental matrix or homography), not pose accuracy. This is clearly stated in Sec. VI-A, but the text should more prominently distinguish this metric from the single-frame pose success rate in Table III to avoid conflation.
  2. [Eq. (11)] The loss balance β in Eq. (11) is not given a value in the implementation details. As a free parameter, β and the adapter training hyperparameters (epochs, learning rate, batch size) should be reported for reproducibility.
  3. [Table I] Typo: 'Y es' should be 'Yes'.
  4. [Throughout] Spacing issue: 'UA V' appears throughout the manuscript; it should be 'UAV'.

Circularity Check

1 steps flagged

Trajectory headline relies on a motion prior generated from the same GNSS reference used for error evaluation, so the 5.88 m/55 km GNSS-denied claim is partly built into the evaluation; adapter and single-frame results remain independent.

specific steps
  1. other [Sec. V Eq. (19); Sec. VI-D Table III; Sec. VII Limitations]
    "To compare trajectory methods under the same realistic, imperfect motion input, we derive a reproducible stochastic odometry prior from consecutive GNSS reference poses T⋆t−1, T⋆t ∈SE(3). We perturb each relative motion by zero-mean noise whose scale increases with translation and rotation, thereby modeling the accumulation of odometry uncertainty without using GNSS as an absolute EKF update: ΔT⋆t = (T⋆t−1)^−1 T⋆t, ˜ΔTt = ΔT⋆t Exp(ξt), ξt ∼ N(0,Σt). GNSS is used only to generate this fixed prior and to score accuracy, never as an absolute localization update."

    The EKF's motion prior is generated from the same GNSS reference poses T⋆ that later define the error metric (Sec. VI-D reports 'mean 3D' and 'success' against the GNSS reference). Eq. 19 injects ground-truth relative increments plus zero-mean noise, so the trajectory is propagated along the reference itself and map updates only correct small synthetic perturbations. The reported 5.88 m mean 3D error over 55 km is therefore not an independent GNSS-denied estimate; it is partly built from the reference used for scoring. The paper acknowledges this in Sec. VII: 'The evaluation is limited to ... a GNSS-reference-derived motion prior.' The abstract, however, presents the same number as evidence for GNSS-denied navigation.

full rationale

The central adapter and single-frame claims are not circular: the adapter is trained on synthetic satellite-derived pairs from UAV-VisLoc and evaluated on the geographically distinct UAV-Terra3D without target training (Fig. 2, Sec. VI-B), and the OSM-filtering ablation holds matcher, DEM lifting, PnP and thresholds fixed (Sec. VI-C). No load-bearing self-citation or imported uniqueness theorem appears; the frozen matcher (RoMav2) is external. However, the trajectory headline (Q4/Table III) is partially circular: the motion prior in Eq. (19) is derived from the same GNSS reference used to compute the error, so the 5.88 m result measures the map updates under an oracle-quality relative motion prior rather than genuine GNSS-denied operation. This closed loop affects the paper's strongest claim, so the overall circularity score is 6. The fix is to re-evaluate with measured onboard odometry (the paper's own future work) or clearly separate the reference-derived prior result from a GNSS-denied claim.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

No new physical entities are invented. The free parameters are hand-set noise and gate values that materially affect the trajectory result. The key non-standard axiom is the GNSS-derived motion prior, which is the main source of circularity.

free parameters (3)
  • EKF motion noise coefficients = α1=0.05, α2=0.01 rad/m, α3=0.05, α4=0.10 m/rad; βp=0.05 m; βR=0.5°
    Hand-selected in Eq. (20). They set the scale of the GNSS-derived motion prior and directly affect EKF gating and the reported trajectory error.
  • χ² gate probability = η=0.05, τ=12.59
    Standard 95% gate for 6-DOF updates, but the choice of gate affects how many visual updates are accepted; no sensitivity analysis is provided.
  • Alignment loss balance β
    Eq. (11) balances log-scale and cosine yaw losses; the value is not reported, affecting adapter training but not directly the final pose.
axioms (5)
  • domain assumption Camera intrinsics and gimbal extrinsics are known and calibrated
    Used in Eqs. (2)-(3) and PnP. The paper states the DJI Air 2S provides no independently validated full 6-DoF ground truth or gimbal calibration suitable for roll/pitch/yaw error evaluation.
  • domain assumption Map layers (NAIP, USGS DEM, OSM) are mutually aligned in a common WGS84/ENU frame after geoid correction
    The entire DEM-lifting and PnP pipeline (Sec. V, Fig. 4) depends on accurate layer alignment and geoid separation; misregistration would directly bias correspondences.
  • domain assumption After OSM filtering, a dominant set of correspondences is terrain-consistent
    The RANSAC-PnP stage requires terrain-consistent matches to form the consensus. The paper itself notes that large unmapped structures can form coherent but incorrect consensus (Sec. IV-B).
  • ad hoc to paper The stochastic odometry model in Eq. (20) with hand-set noise represents a realistic relative-motion prior
    The prior is generated from GNSS reference poses rather than an IMU or visual odometry stream. It is not a valid stand-in for GNSS-denied odometry and inflates the trajectory result.
  • domain assumption Synthetic satellite-derived training pairs transfer to real UAV imagery
    The adapter is trained only on synthetic UAV-VisLoc pairs (China) and evaluated on real UAV-Terra3D (Austin). Evidence is limited to eight local sites and mostly nadir views, as acknowledged in the limitations.

pith-pipeline@v1.3.0-alltime-deepseek · 11573 in / 10267 out tokens · 92579 ms · 2026-08-02T03:01:00.322388+00:00 · methodology

0 comments
read the original abstract

We present AeroMap3D, a monocular 6-DoF UAV localization system that anchors onboard imagery to visual, geometric, and semantic map priors for GNSS-denied navigation. AeroMap3D addresses two fundamental challenges in map-referenced aerial localization: the cross-view discrepancy between UAV imagery and satellite maps, and the structural inconsistency between bare-earth digital elevation models (DEMs) and urban scenes. First, we introduce a lightweight adapter that enables a dense matcher pretrained on internet-scale generic data to perform reliable UAV-to-map registration without finetuning. By estimating the scale ratio and yaw offset between the UAV image and map tile, the adapter removes the dominant geometric misalignment induced by altitude, camera field of view, and heading before dense correspondence estimation. Second, AeroMap3D lifts 2D UAV-map correspondences onto DEM terrain while using OpenStreetMap annotations to reject semantically unreliable matches before RANSAC-PnP pose estimation, thereby reducing errors caused by unmodeled building heights and off-nadir structures. Delayed map-based pose measurements are further fused with relative-motion priors using a delayed-state EKF for continuous trajectory estimation. Without UAV-Terra3D retraining or tuning, AeroMap3D localizes all trajectories across eight Austin sites within 50 m and achieves 5.88 m mean 3D error over 55 km of flight.

Figures

Figures reproduced from arXiv: 2607.14009 by Luis Sentis, Zhiyun Deng.

Figure 1
Figure 1. Figure 1: Overview of AeroMap3D. Compared with prior 6- DoF UAV localization systems that rely on textured 3D maps, DSMs, or render databases, AeroMap3D uses widely available satellite imagery, bare-earth DEM, and OSM layers, enabling better generalization to unexplored areas. A scale/yaw adapter and OSM masking condition UAV–map correspondences be￾fore DEM lifting and standard RANSAC–PnP. satellite imagery [11], ba… view at source ↗
Figure 2
Figure 2. Figure 2: Synthetic training data for the scale–yaw adapter. Training UAV–map pairs are generated by applying known transformations to satellite images, providing direct supervi￾sion without real UAV images or manual labeling. and the resulting reprojection residual is no longer a random matching error but a structured geometric bias. Robust Pose Objective. AeroMap3D first removes known structural regions in the map… view at source ↗
Figure 3
Figure 3. Figure 3: AeroMap3D localization pipeline. Starting from a coarse map initialization, obtained by retrieval in prior work or by an operator in our experiments, the system propagates with a relative-motion prior. A scale/yaw adapter enables dense UAV–map matching; OSM masking and DEM lifting form semantically reliable 2D–3D correspondences; RANSAC–PnP estimates the global camera pose; and a timestamp-aligned EKF fuse… view at source ↗
Figure 4
Figure 4. Figure 4: UAV-Terra3D map priors. Eight Austin sites combine WGS84-aligned NAIP imagery [11] for registration (top; 1 m/pixel), OpenStreetMap [13] for semantic filtering (middle), and USGS DEMs [12] for 2D–3D lifting (bottom; 1 m/pixel) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cross-domain UAV-to-map registration and run￾time. Left: Direct-matcher success (bars; I-shaped error bars indicate the minimum–maximum range across the dataset) versus adapter-enhanced success (line). Our adapter raises RoMav2 from 62.4% to 99.2% and Tiny-RoMa from 42.1% to 92.5%. Right: Jetson runtimes (bars, including 35 ms adapter overhead) and RTX 5090 runtimes (dashed line); RoMav2 exceeds the Jetson… view at source ↗
Figure 7
Figure 7. Figure 7: Representative 2-km UAV-Terra3D trajectory. Top: map-space trajectories. Bottom: altitude profiles. AeroMap3D closely follows the ground truth while map-unanchored odom￾etry and SLAM drift; GeoVINS bounds drift but remains less accurate than metric PnP updates. The delayed-update EKF accommodates these asynchronous visual updates. C. Q3: Single-Frame Camera Pose Estimation The upper half of Table III shows… view at source ↗
Figure 6
Figure 6. Figure 6: Long-horizon relative drift. RPE versus segment length; lines show medians, shaded regions show the 25–75% and 5–95% ranges, and faint traces show individual trajec￾tories. AeroMap3D remains below 6 m RPE at 2 km, while reference odometry, ORB-VO, and ORB-SLAM2 accumulate substantial drift. every evaluated matcher. In particular, Adapter+RoMav2 reaches 99.2%, a 36.8-point gain, providing the geometrically … 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

34 extracted references · 1 canonical work pages

  1. [1]

    Visual odometry

    David Nistér, Oleg Naroditsky, and James Bergen. Visual odometry. InProceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, volume 1, pages I–652–I–659, 2004

    2004

  2. [2]

    Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras.IEEE transactions on robotics, 33(5):1255– 1262, 2017

    Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras.IEEE transactions on robotics, 33(5):1255– 1262, 2017

    2017

  3. [3]

    Netvlad: Cnn architecture for weakly supervised place recognition

    Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5297– 5307, 2016

    2016

  4. [4]

    Geovins: Geographic-visual-inertial navigation system for large- scale drift-free aerial state estimation.IEEE Transactions on Robotics, 2025

    Chunyu Li, Mengfan He, Chao Chen, Jiacheng Liu, Xu Lyu, Guoquan Huang, and Ziyang Meng. Geovins: Geographic-visual-inertial navigation system for large- scale drift-free aerial state estimation.IEEE Transactions on Robotics, 2025

    2025

  5. [5]

    Scaling image geo-localization to continent level

    Philipp Lindenberger, Paul-Edouard Sarlin, Jan Hosang, Marc Pollefeys, Simon Lynen, and Eduard Trulls. Scaling image geo-localization to continent level. Advances in Neural Information Processing Systems, 38:134057–134092, 2026

    2026

  6. [6]

    Understanding pnp - solvepnp

    OpenCV Forum. Understanding pnp - solvepnp. https://forum.opencv.org/t/ understanding-pnp-solvepnp/7193, 2022

    2022

  7. [7]

    Real- time geo-localization using satellite imagery and topography for unmanned aerial vehicles

    Shuxiao Chen, Xiangyu Wu, Mark W Mueller, and Koushil Sreenath. Real- time geo-localization using satellite imagery and topography for unmanned aerial vehicles. In2021 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 2275–2281. IEEE, 2021

    2021

  8. [8]

    Uavd4l: A large-scale dataset for uav 6-dof localization

    Rouwan Wu, Xiaoya Cheng, Juelin Zhu, Xuxiang Liu, Maojun Zhang, and Shen Yan. Uavd4l: A large-scale dataset for uav 6-dof localization. In2024 International Conference on 3D Vision (3DV), pages 1574–1583. IEEE, 2024

    2024

  9. [9]

    Pilot: Neural pixel-to-3d registration for uav-based ego and target geo-localization

    Xiaoya Cheng, Long Wang, Yan Liu, Xinyi Liu, Hanlin Tan, Yu Liu, Maojun Zhang, and Shen Yan. Pilot: Neural pixel-to-3d registration for uav-based ego and target geo-localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5379–5388, 2026

    2026

  10. [10]

    Ortholoc: Uav 6-dof localization and calibration using orthographic geodata

    Oussema Dhaouadi, Riccardo Marin, Johannes Meier, Jacques Kaiser, and Daniel Cremers. Ortholoc: Uav 6-dof localization and calibration using orthographic geodata. InAdvances in Neural Information Processing Systems, 2025

    2025

  11. [11]

    Department of Agriculture

    Farm Service Agency U.S. Department of Agriculture. National agriculture imagery program (naip). https://naip-usdaonline.hub.arcgis.com/, 2026

    2026

  12. [12]

    Geological Survey

    U.S. Geological Survey. 3d elevation program (3dep). https://www.usgs.gov/ 3d-elevation-program, 2026

    2026

  13. [13]

    Planet dump retrieved from https://planet.osm.org

    OpenStreetMap contributors. Planet dump retrieved from https://planet.osm.org . https://www.openstreetmap.org, 2026

    2026

  14. [14]

    Loftr: Detector-free local feature matching with transformers

    Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922– 8931, 2021

    2021

  15. [15]

    Roma: Robust dense feature matching

    Johan Edstedt, Qiyu Sun, Georg Bökman, Mårten Wadenbäck, and Michael Felsberg. Roma: Robust dense feature matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790–19800, 2024

    2024

  16. [16]

    Roma v2: Harder better faster denser feature matching.arXiv preprint arXiv:2511.15706, 2025

    Johan Edstedt, David Nordström, Yushan Zhang, Georg Bökman, Jonathan Aster- mark, Viktor Larsson, Anders Heyden, Fredrik Kahl, Mårten Wadenbäck, and Michael Felsberg. Roma v2: Harder better faster denser feature matching.arXiv preprint arXiv:2511.15706, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Magnetometer calibration for small unmanned aerial vehicles using cooperative flight data.Sensors, 20(2):538, 2020

    Roberto Opromolla. Magnetometer calibration for small unmanned aerial vehicles using cooperative flight data.Sensors, 20(2):538, 2020. doi: 10.3390/s20020538

    work page doi:10.3390/s20020538 2020

  18. [18]

    Modelling and calibration of time- varying magnetometer biases in small unmanned aerial vehicles

    Andrew Stewart and Demoz Gebre-Egziabher. Modelling and calibration of time- varying magnetometer biases in small unmanned aerial vehicles. InProceedings of the 28th International Technical Meeting of the Satellite Division of the Institute of Navigation (ION GNSS+ 2015), pages 777–783, Tampa, Florida, 2015

    2015

  19. [19]

    Optimal randomized ransac.IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(8):1472–1482, 2008

    Ond ˇrej Chum and Ji ˇrí Matas. Optimal randomized ransac.IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(8):1472–1482, 2008

    2008

  20. [20]

    Alto: A large-scale dataset for uav visual place recognition and localization.arXiv preprint arXiv:2207.12317, 2022

    Ivan Cisneros, Peng Yin, Ji Zhang, Howie Choset, and Sebastian Scherer. Alto: A large-scale dataset for uav visual place recognition and localization.arXiv preprint arXiv:2207.12317, 2022

    Pith/arXiv arXiv 2022

  21. [21]

    Uav-visloc: A large-scale dataset for uav visual localization.arXiv preprint arXiv:2405.11936, 2024

    Wenjia Xu, Yaxuan Yao, Jiaqi Cao, Zhiwei Wei, Chunbo Liu, Jiuniu Wang, and Mugen Peng. Uav-visloc: A large-scale dataset for uav visual localization.arXiv preprint arXiv:2405.11936, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Aerialvl: A dataset, baseline and algorithm framework for aerial-based visual localization with reference map.IEEE Robotics and Automation Letters, 9(10):8210–8217, 2024

    Mengfan He, Chao Chen, Jiacheng Liu, Chunyu Li, Xu Lyu, Guoquan Huang, and Ziyang Meng. Aerialvl: A dataset, baseline and algorithm framework for aerial-based visual localization with reference map.IEEE Robotics and Automation Letters, 9(10):8210–8217, 2024

    2024

  23. [23]

    Orthotrack: Continuous 6-dof uav trajectory estimation anchored in public orthophotos.arXiv preprint arXiv:2606.25245, 2026

    Oussema Dhaouadi, Zuria Bauer, Johannes Michael Meier, Olaf Wysocki, Marc Pollefeys, and Daniel Cremers. Orthotrack: Continuous 6-dof uav trajectory estimation anchored in public orthophotos.arXiv preprint arXiv:2606.25245, 2026

    Pith/arXiv arXiv 2026

  24. [24]

    Dinov2-based uav visual self-localization in low-altitude urban environments.IEEE Robotics and Automation Letters, 2025

    Jiaqiang Yang, Danyang Qin, Huapeng Tang, Sili Tao, Haoze Bie, and Lin Ma. Dinov2-based uav visual self-localization in low-altitude urban environments.IEEE Robotics and Automation Letters, 2025

    2025

  25. [25]

    Selavpr++: Towards seamless adaptation of foundation models for efficient place recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Feng Lu, Tong Jin, Xiangyuan Lan, Lijun Zhang, Yunpeng Liu, Yaowei Wang, and Chun Yuan. Selavpr++: Towards seamless adaptation of foundation models for efficient place recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  26. [26]

    Effovpr: Effective foundation model utilization for visual place recognition

    Issar Tzachor, Boaz Lerner, Matan Levy, Michael Green, Tal Berkovitz Shalev, Gavriel Habib, Dvir Samuel, Noam Zailer, Or Shimshi, Nir Darshan, et al. Effovpr: Effective foundation model utilization for visual place recognition. InInternational Conference on Learning Representations, volume 2025, pages 42817–42839, 2025

    2025

  27. [27]

    Efficient visual place recognition through multimodal semantic knowledge integration

    Sitao Zhang, Hongda Mao, Qingshuang Chen, and Yelin Kim. Efficient visual place recognition through multimodal semantic knowledge integration. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5601–5610, 2025

    2025

  28. [28]

    Superglue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabi- novich. Superglue: Learning feature matching with graph neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020

    2020

  29. [29]

    Le, and Hartwig Adam

    Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V . Le, and Hartwig Adam. Searching for mobilenetv3. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1314–1324, 2019

    2019

  30. [30]

    MIT Press, 2005

    Sebastian Thrun, Wolfram Burgard, and Dieter Fox.Probabilistic Robotics. MIT Press, 2005

    2005

  31. [31]

    Orb: An efficient alternative to sift or surf

    Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In2011 International conference on computer vision, pages 2564–2571. Ieee, 2011

    2011

  32. [32]

    Lightglue: Local feature matching at light speed

    Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. InProceedings of the IEEE/CVF international conference on computer vision, pages 17627–17638, 2023

    2023

  33. [33]

    Superpoint: Self- supervised interest point detection and description

    Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self- supervised interest point detection and description. InProceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224–236, 2018

    2018

  34. [34]

    Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El- Nouby, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023