pith. sign in

arxiv: 2606.25245 · v2 · pith:NDQNAHVBnew · submitted 2026-06-24 · 💻 cs.CV

OrthoTrack: Continuous 6-DoF UAV Trajectory Estimation Anchored in Public Orthophotos

Oussema Dhaouadi , Zuria Bauer , Johannes Michael Meier , Olaf Wysocki , Marc Pollefeys , Daniel Cremers This is my paper

Pith reviewed 2026-06-30 09:44 UTC · model grok-4.3

classification 💻 cs.CV
keywords UAV trajectory estimationorthophoto matching6-DoF pose estimationvisual odometrygeolocalizationoptical flowpublic map priorscontinuous pose
0
0 comments X

The pith

OrthoTrack produces continuous metric 6-DoF UAV trajectories by anchoring keyframes to public orthophotos and propagating via optical flow.

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

The paper introduces OrthoTrack, a training-free pipeline that estimates absolute 6-DoF poses for UAVs at every frame using only publicly available orthophotos and surface models. It selects keyframes, matches them directly to the orthophoto, lifts the 2D matches to metric 3D positions with the surface model, and then uses optical flow to carry those anchored correspondences forward to all intermediate frames. This produces drift-free, metrically scaled trajectories in real time without GPS, post-processing alignment, or learned models. The method is demonstrated on both a new large-scale synthetic benchmark and real-world sequences, where it exceeds prior visual odometry and single-frame localization baselines.

Core claim

OrthoTrack matches keyframes against the orthophoto and lifts correspondences to metric 3D via the surface model. It then propagates these map-anchored correspondences to intermediate frames with optical flow, producing absolute, metrically scaled poses at every frame without GPS or post-hoc alignment.

What carries the argument

Keyframe-to-orthophoto matching followed by 3D lifting via the surface model and optical-flow propagation to non-key frames.

If this is right

  • Absolute metric scale is available at every frame without any post-hoc correction.
  • The system runs in real time on a single GPU using only public map data.
  • New regions can be handled without site-specific training or adaptation.
  • Continuous trajectories are produced even though only keyframes are matched to the map.
  • The approach outperforms both relative visual odometry and single-frame geo-localization baselines.

Where Pith is reading between the lines

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

  • The same anchoring idea could be tested on satellite imagery or other public elevation sources if the matching step remains reliable.
  • Integration with onboard inertial sensors might further improve robustness when visual matches are sparse.
  • The pipeline could support repeated flights over the same area by reusing the same public map across missions.
  • Performance on outdated maps would reveal how much temporal mismatch the matching stage can tolerate before drift appears.

Load-bearing premise

Orthophoto-to-image matching must yield sufficiently accurate and repeatable correspondences that survive viewpoint, lighting, and temporal differences.

What would settle it

Trajectory error remains low on sequences where the orthophoto and UAV imagery differ substantially in season, time of day, or construction changes.

Figures

Figures reproduced from arXiv: 2606.25245 by Daniel Cremers, Johannes Michael Meier, Marc Pollefeys, Olaf Wysocki, Oussema Dhaouadi, Zuria Bauer.

Figure 1
Figure 1. Figure 1: OrthoTrack Estimates Continuous 6-DoF UAV Trajectories From [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: OrthoTrack Pipeline. Keyframe localization (top): a visibility-aware crop selects DOP D k and DSM S k from the previous pose. The dense matcher M matches keyframe Ik against D k , correspondences are lifted to 3D via S k , and PnP-RANSAC re￾covers the keyframe pose Tk. Inter-frame tracking (bottom): optical flow O propagates keyframe-anchored correspondences from previous frame It−1 to current frame It for… view at source ↗
Figure 3
Figure 3. Figure 3: Adaptive Reprojection Thresholds. The absolute threshold (blue) starts relaxed during warmup, then settles at ebase. The relative threshold (red) tightens over D frames. Rapid drift is caught by the absolute bound, slow drift by the tightening relative bound. maining points cluster in a small image region, making PnP ill-conditioned), or (4) reprojection error et exceeds an adaptive threshold. A fixed repr… view at source ↗
Figure 4
Figure 4. Figure 4: MovingDrone Locations. 194 sequences across 21 Berlin regions. Insets show representative trajectories over satellite imagery. The heatmap encodes sequence density. sequences anywhere within the city’s coverage at negligible cost, circumventing the airspace restrictions, privacy regulations, and expensive RTK-GNSS super￾vision that make real UAV data collection impractical at metropolitan scale. Be￾cause t… view at source ↗
Figure 5
Figure 5. Figure 5: MovingDrone Multi-Modal Data Per Frame. Each rendered UAV frame is paired with depth, surface normals from the mesh, LiDAR, multi-year DOPs, a DSM, and LoD1/2 building meshes. Comparing the three DOPs reveals the temporal domain gap: lighting, vegetation, and even building footprints change across years. bining GPS drift with motor vibration, (ii) stochastic wind-gust events, and (iii) velocity-dependent d… view at source ↗
Figure 6
Figure 6. Figure 6: Trajectory Comparison. VO/SLAM baselines receive oracle Sim(3) align￾ment yet drift or collapse on long flights. Piecewise-anchored DROID-SLAM (pw) re￾duces drift but accumulates errors between anchors. OrthoTrack produces absolute poses without alignment and remains accurate. 5.2 Comparison with State of the Art Tab. 2 presents results on MovingDrone and on real-world UAVScenes [47]. Or￾thoTrack is the on… view at source ↗
Figure 7
Figure 7. Figure 7: Per-Frame Localization Quality. LoD2 building wireframes (red) overlaid on the orthophoto are shifted by each method’s position error, making misalignment directly visible. priors, which are often unavailable or unreliable in real-world aerial scenarios. DPVO [40] and ORB-SLAM3 [3] exhibit the same degeneracy, confirming a fun￾damental limitation of relative-pose methods on aerial scenes. Moreover, DROID r… view at source ↗
read the original abstract

Continuous 6-DoF pose estimation is essential for autonomous UAV operations. Yet, existing visual odometry and SLAM methods accumulate drift and yield only relative, up-to-scale trajectories. Single-frame geo-localization, in turn, discards temporal continuity and remains too slow for real-time use. We present OrthoTrack, a training-free system that estimates continuous 6-DoF UAV trajectories using only publicly available orthophotos and surface models as a map prior. OrthoTrack matches keyframes against the orthophoto and lifts correspondences to metric 3D via the surface model. It then propagates these map-anchored correspondences to intermediate frames with optical flow, producing absolute, metrically scaled poses at every frame without GPS or post-hoc alignment. We also introduce the MovingDrone Dataset, a large-scale benchmark pairing photorealistic UAV sequences with dense 6-DoF ground truth and co-registered multi-modal geodata including multi-temporal orthophotos. On MovingDrone and real-world benchmarks, OrthoTrack runs in real time on a single GPU. It outperforms all baselines by a large margin, even those receiving oracle scale and alignment. By relying on publicly available geodata, OrthoTrack enables deployment to new regions without site-specific adaptation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The paper presents OrthoTrack, a training-free pipeline for continuous 6-DoF UAV trajectory estimation that anchors keyframes to public orthophotos, lifts 2D-2D correspondences to metric 3D using a surface model, and propagates the anchored points to all frames via optical flow to obtain absolute, metrically scaled poses at every timestep without GPS or post-hoc alignment. It introduces the MovingDrone Dataset pairing photorealistic UAV sequences with dense ground truth and co-registered multi-temporal geodata. Experiments claim real-time GPU operation and large-margin outperformance over VO/SLAM and geo-localization baselines, including those given oracle scale and alignment.

Significance. If the performance and robustness claims hold, the work would enable practical, drift-free absolute localization for UAVs in new regions using only publicly available orthophotos and DSMs, removing the need for site-specific training or GPS. The training-free design, real-time capability, and introduction of the MovingDrone Dataset (with its multi-modal, multi-temporal registration) are clear strengths that could support reproducible benchmarking in aerial visual localization.

major comments (3)
  1. [§3.2] §3.2 (Keyframe-to-orthophoto matching): The module is described only at the level of 'matches keyframes against the orthophoto' with no specification of the descriptor, similarity metric, RANSAC threshold, or inlier selection criteria. This is load-bearing for the central claim because the accuracy of 3D lifting via the surface model and the drift-free property of the subsequent optical-flow propagation both rest on the repeatability and precision of these correspondences under nadir-to-oblique viewpoint change, illumination variation, and possible multi-temporal map misalignment.
  2. [Table 2] Table 2 (MovingDrone quantitative results): The reported ATE and RPE values show large-margin gains, yet the table omits per-sequence inlier rates for the orthophoto matching step, 3D lifting error statistics, and the fraction of frames where matching fails. Without these, it is impossible to verify whether the absolute metric property is maintained or whether the outperformance is driven by easy sequences where matching succeeds.
  3. [§4.3] §4.3 (Real-world benchmarks): The evaluation does not report any controlled stress test on temporal map misalignment or strong lighting changes, both of which are explicitly flagged as failure modes for orthophoto-to-image matching. This directly affects the claim that the method works 'even those receiving oracle scale and alignment' across diverse conditions.
minor comments (3)
  1. [Figure 2] Figure 2: The pipeline diagram would be clearer if the optical-flow propagation arrows were labeled with the frame indices they connect.
  2. [§2] §2 (Related work): Several recent learned geo-localization methods that also use orthophoto priors are not cited; adding them would better situate the training-free design choice.
  3. [Eq. (3)] Eq. (3): The notation for the lifted 3D point P does not explicitly state whether the surface model is assumed to be in the same coordinate frame as the orthophoto or requires an additional rigid transform.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight areas where additional detail and analysis will strengthen the manuscript. We address each major comment below and will revise the paper accordingly.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Keyframe-to-orthophoto matching): The module is described only at the level of 'matches keyframes against the orthophoto' with no specification of the descriptor, similarity metric, RANSAC threshold, or inlier selection criteria. This is load-bearing for the central claim because the accuracy of 3D lifting via the surface model and the drift-free property of the subsequent optical-flow propagation both rest on the repeatability and precision of these correspondences under nadir-to-oblique viewpoint change, illumination variation, and possible multi-temporal map misalignment.

    Authors: We agree that the description in §3.2 lacks the necessary implementation specifics for reproducibility. In the revised manuscript we will expand this section to explicitly state the feature descriptor, similarity metric, RANSAC threshold, and inlier selection criteria employed, thereby clarifying how the method maintains correspondence quality under the cited viewpoint, illumination, and map-misalignment conditions. revision: yes

  2. Referee: [Table 2] Table 2 (MovingDrone quantitative results): The reported ATE and RPE values show large-margin gains, yet the table omits per-sequence inlier rates for the orthophoto matching step, 3D lifting error statistics, and the fraction of frames where matching fails. Without these, it is impossible to verify whether the absolute metric property is maintained or whether the outperformance is driven by easy sequences where matching succeeds.

    Authors: We concur that these per-sequence diagnostics are important for validating the absolute-metric claim. We will revise Table 2 (or introduce a companion table) to report per-sequence inlier rates, 3D lifting error statistics, and the fraction of frames where orthophoto matching fails, providing a clearer picture of performance variation across the benchmark. revision: yes

  3. Referee: [§4.3] §4.3 (Real-world benchmarks): The evaluation does not report any controlled stress test on temporal map misalignment or strong lighting changes, both of which are explicitly flagged as failure modes for orthophoto-to-image matching. This directly affects the claim that the method works 'even those receiving oracle scale and alignment' across diverse conditions.

    Authors: The referee correctly notes the absence of targeted stress tests on these failure modes. Although the MovingDrone dataset already includes multi-temporal orthophotos, we will add controlled experiments and analysis in §4.3 that isolate temporal misalignment and strong lighting variations, directly supporting the robustness claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper describes a training-free pipeline that matches keyframes to public orthophotos, lifts correspondences via surface models, and propagates via optical flow to produce absolute poses. No equations, fitted parameters, or self-citations are presented that would make any claimed output equivalent to its inputs by construction. The central claims rest on an externally introduced benchmark dataset (MovingDrone) with independent ground truth and comparisons to baselines, rendering the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no identifiable free parameters, axioms, or invented entities; the method description does not introduce new mathematical constructs or fitted constants.

pith-pipeline@v0.9.1-grok · 5774 in / 1283 out tokens · 32519 ms · 2026-06-30T09:44:14.053334+00:00 · methodology

Review history (2 revisions) →

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

61 extracted references · 10 canonical work pages · 4 internal anchors

  1. [1]

    Bochkovskiy, A., Delaunoy, A., Germain, H., Santos, M., Zhou, Y., Richter, S., Koltun, V.: Depth pro: Sharp monocular metric depth in less than a second. In: Int. Conf. Learn. Represent. (2025)

    2025

  2. [2]

    In: IEEE Conf

    Cabon, Y., Stoffl, L., Antsfeld, L., Csurka, G., Chidlovskii, B., Revaud, J., Leroy, V.: Must3r: Multi-view network for stereo 3d reconstruction. In: IEEE Conf. Com- put. Vis. Pattern Recog. pp. 1050–1060 (2025)

    2025

  3. [3]

    IEEE Trans

    Campos, C., Elvira, R., Gómez, J.J., Montiel, J.M.M., Tardós, J.D.: ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Trans. Robot.37(6), 1874–1890 (2021)

    2021

  4. [4]

    In: IEEE Conf

    Chen, G., Fu, T., Chen, H., Teng, W., Xiao, H., Zhao, Y.: Rdd: Robust feature detector and descriptor using deformable transformer. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6394–6403 (2025) 16 O. Dhaouadi et al

    2025

  5. [5]

    arXiv preprint arXiv:2207.12317 (2022)

    Cisneros, I., Yin, P., Zhang, J., Choset, H., Scherer, S.: Alto: A large-scale dataset for uav visual place recognition and localization. arXiv preprint arXiv:2207.12317 (2022)

    work page arXiv 2022

  6. [6]

    Dhaouadi,O.,Marin,R.,Meier,J.M.,Kaiser,J.,Cremers,D.:Ortholoc:UAV6-dof localization and calibration using orthographic geodata. In: Adv. Neural Inform. Process. Syst. (2025)

    2025

  7. [7]

    In: IEEE Conf

    Edstedt, J., Athanasiadis, I., Wadenbäck, M., Felsberg, M.: Dkm: Dense kernelized feature matching for geometry estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17765–17775 (2023)

    2023

  8. [8]

    arXiv preprint arXiv:2503.07347 (2025)

    Edstedt, J., Bökman, G., Wadenbäck, M., Felsberg, M.: Dad: Distilled reinforce- ment learning for diverse keypoint detection. arXiv preprint arXiv:2503.07347 (2025)

    work page arXiv 2025

  9. [9]

    Edstedt, J., Bökman, G., Wadenbäck, M., Felsberg, M.: DeDoDe: Detect, Don’t Describe — Describe, Don’t Detect for Local Feature Matching. In: Int. Conf. 3D Vis. IEEE (2024)

    2024

  10. [10]

    arXiv preprint arXiv:2511.15706 , year=

    Edstedt, J., Nordström, D., Zhang, Y., Bökman, G., Astermark, J., Larsson, V., Heyden, A., Kahl, F., Wadenbäck, M., Felsberg, M.: Roma v2: Harder better faster denser feature matching. arXiv preprint arXiv:2511.15706 (2025)

    work page arXiv 2025

  11. [11]

    In: IEEE Conf

    Edstedt, J., Sun, Q., Bökman, G., Wadenbäck, M., Felsberg, M.: Roma: Robust dense feature matching. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19790– 19800 (2024)

    2024

  12. [12]

    Google: Google earth studio.https://earth.google.com/studio/(2026)

    2026

  13. [13]

    Scientific reports6(1), 22574 (2016)

    Hodgson, J.C., Baylis, S.M., Mott, R., Herrod, A., Clarke, R.H.: Precision wildlife monitoring using unmanned aerial vehicles. Scientific reports6(1), 22574 (2016)

    2016

  14. [14]

    IEEE Trans

    Hu, M., Yin, W., Zhang, C., Cai, Z., Long, X., Chen, H., Wang, K., Yu, G., Shen, C., Shen, S.: Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Trans. Pattern Anal. Mach. Intell.46(12), 10579–10596 (2024)

    2024

  15. [15]

    In: IEEE Conf

    Jang, W., Weinzaepfel, P., Leroy, V., Agapito, L., Revaud, J.: Pow3r: Empowering unconstrained 3d reconstruction with camera and scene priors. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1071–1081 (2025)

    2025

  16. [16]

    In: AAAI

    Ji, Y., He, B., Tan, Z., Wu, L.: Game4loc: A uav geo-localization benchmark from game data. In: AAAI. vol. 39, pp. 3913–3921 (2025)

    2025

  17. [17]

    IET Radar, Sonar & Nav- igation12(2), 151–164 (2018)

    Jordan, S., Moore, J., Hovet, S., Box, J., Perry, J., Kirsche, K., Lewis, D., Tse, Z.T.H.: State-of-the-art technologies for uav inspections. IET Radar, Sonar & Nav- igation12(2), 151–164 (2018)

    2018

  18. [18]

    IEEE Trans

    Ke, B., Qu, K., Wang, T., Metzger, N., Huang, S., Li, B., Obukhov, A., Schindler, K.: Marigold: Affordable adaptation of diffusion-based image generators for image analysis. IEEE Trans. Pattern Anal. Mach. Intell. (2025)

    2025

  19. [19]

    MapAnything: Universal Feed-Forward Metric 3D Reconstruction

    Keetha, N., Müller, N., Schönberger, J., Porzi, L., Zhang, Y., Fischer, T., Knapitsch, A., Zauss, D., Weber, E., Antunes, N., et al.: Mapanything: Univer- sal feed-forward metric 3d reconstruction. arXiv preprint arXiv:2509.13414 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  20. [20]

    Künzel, J., Hilsmann, A., Eisert, P.: Ripe: Reinforcement learning on unlabeled image pairs for robust keypoint extraction. In: Int. Conf. Comput. Vis. pp. 4868– 4877 (2025)

    2025

  21. [21]

    Online geodata platform (2025), https://data.geobasis-bb.de, data licensed under Datenlizenz Deutschland Na- mensnennung 2.0

    Landesvermessung und Geobasisinformation Brandenburg (LGB): Opengeo- data.brandenburg - data portal of geobasis-de. Online geodata platform (2025), https://data.geobasis-bb.de, data licensed under Datenlizenz Deutschland Na- mensnennung 2.0

    2025

  22. [22]

    Leroy, V., Cabon, Y., Revaud, J.: Grounding image matching in 3d with mast3r. In: Eur. Conf. Comput. Vis. pp. 71–91. Springer (2024) OrthoTrack 17

    2024

  23. [23]

    Li, X., Rao, T., Pan, C.: Edm: Efficient deep feature matching. In: Int. Conf. Comput. Vis. pp. 26198–26208 (2025)

    2025

  24. [24]

    Liang, Y., Hu, Y., Shao, W., Fu, Y.: Learning dense feature matching via lifting single 2d image to 3d space. In: Int. Conf. Comput. Vis. pp. 6621–6631 (2025)

    2025

  25. [25]

    Depth Anything 3: Recovering the Visual Space from Any Views

    Lin, H., Chen, S., Liew, J., Chen, D.Y., Li, Z., Shi, G., Feng, J., Kang, B.: Depth anything 3: Recovering the visual space from any views. arXiv preprint arXiv:2511.10647 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [26]

    Lindenberger, P., Sarlin, P.E., Pollefeys, M.: Lightglue: Local feature matching at light speed. In: Int. Conf. Comput. Vis. pp. 17627–17638 (2023)

    2023

  27. [27]

    In: IEEE Int

    Liu, Y., Lai, W., Zhao, Z., Xiong, Y., Zhu, J., Cheng, J., Xu, Y.: Liftfeat: 3d geometry-aware local feature matching. In: IEEE Int. Conf. Robot. Autom. pp. 11714–11720. IEEE (2025)

    2025

  28. [28]

    In: AAAI

    Lucas, B.D., Kanade, T.: An iterative image registration technique with an appli- cation to stereo vision. In: AAAI. vol. 2, pp. 674–679 (1981)

    1981

  29. [29]

    Pattern Anal

    Nistér,D.:Anefficientsolutiontothefive-pointrelativeposeproblem.IEEETrans. Pattern Anal. Mach. Intell.26(6), 756–770 (2004)

    2004

  30. [30]

    Panek, V., Kukelova, Z., Sattler, T.: Meshloc: Mesh-based visual localization. In: Eur. Conf. Comput. Vis. pp. 589–609. Springer (2022)

    2022

  31. [31]

    In: IEEE Conf

    Piccinelli, L., Yang, Y.H., Sakaridis, C., Segu, M., Li, S., Van Gool, L., Yu, F.: Unidepth: Universal monocular metric depth estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10106–10116 (2024)

    2024

  32. [32]

    In: IEEE Conf

    Potje, G., Cadar, F., Araujo, A., Martins, R., Nascimento, E.R.: Xfeat: Accelerated features for lightweight image matching. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2682–2691 (2024)

    2024

  33. [33]

    Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: Superglue: Learning featurematchingwithgraphneuralnetworks.In:IEEEConf.Comput.Vis.Pattern Recog. pp. 4938–4947 (2020)

    2020

  34. [34]

    In: Proceedings of the first workshop on micro aerial vehicle networks, systems, and applications for civilian use

    Scherer, J., Yahyanejad, S., Hayat, S., Yanmaz, E., Andre, T., Khan, A., Vukadi- novic, V., Bettstetter, C., Hellwagner, H., Rinner, B.: An autonomous multi-uav system for search and rescue. In: Proceedings of the first workshop on micro aerial vehicle networks, systems, and applications for civilian use. pp. 33–38 (2015)

    2015

  35. [35]

    arXiv preprint arXiv:2205.11567 (2022)

    Schleiss, M., Rouatbi, F., Cremers, D.: Vpair–aerial visual place recognition and localization in large-scale outdoor environments. arXiv preprint arXiv:2205.11567 (2022)

    work page arXiv 2022

  36. [36]

    Senate Department for Urban Development, Building and Housing (Senatsverwal- tung für Stadtentwicklung, Bauen und Wohnen), Berlin: Geoportal berlin (geospa- tial data infrastructure berlin).https://gdi.berlin.de(2026), accessed: 2026- 02-24

    2026

  37. [37]

    arXiv e-prints pp

    Sun, J., Liu, K., Zhang, C., Chen, C., Shen, J., Vong, C.M.: Cross-view uav geo- localization with precision-focused efficient design: A hierarchical distillation ap- proach with multi-view refinement. arXiv e-prints pp. arXiv–2510 (2025)

    2025

  38. [38]

    Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Eur. Conf. Comput. Vis. pp. 402–419. Springer (2020)

    2020

  39. [39]

    Teed, Z., Deng, J.: Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. In: Adv. Neural Inform. Process. Syst. vol. 34, pp. 16558–16569 (2021)

    2021

  40. [40]

    Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. In: Adv. Neural Inform. Process. Syst. vol. 36 (2024)

    2024

  41. [41]

    Terzakis, G., Lourakis, M.: A consistently fast and globally optimal solution to the perspective-n-point problem. In: Eur. Conf. Comput. Vis. pp. 478–494. Springer (2020) 18 O. Dhaouadi et al

    2020

  42. [42]

    virtualcitymap

    virtualcitymap berlin 3d city model.https : / / berlin . virtualcitymap . de/ (2025),https://berlin.virtualcitymap.de/

    2025

  43. [43]

    In: IEEE Conf

    Vuong, K., Ghosh, A., Ramanan, D., Narasimhan, S., Tulsiani, S.: Aerialmegadepth: Learning aerial-ground reconstruction and view synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21674–21684 (2025)

    2025

  44. [44]

    In: IEEE Conf

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Vi- sual geometry grounded transformer. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5294–5306 (2025)

    2025

  45. [45]

    Wang, R., Xu, S., Dong, Y., Deng, Y., Xiang, J., Lv, Z., Sun, G., Tong, X., Yang, J.: Moge-2: Accurate monocular geometry with metric scale and sharp details. In: Adv. Neural Inform. Process. Syst. (2025)

    2025

  46. [46]

    In: IEEE Conf

    Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: Dust3r: Geometric 3d vision made easy. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 20697–20709 (2024)

    2024

  47. [47]

    Wang, S., Li, S., Zhang, Y., Yu, S., Yuan, S., She, R., Guo, Q., Zheng, J., Howe, O.K., Chandra, L., et al.: Uavscenes: A multi-modal dataset for uavs. In: Int. Conf. Comput. Vis. pp. 28946–28958 (2025)

    2025

  48. [48]

    In: IEEE Conf

    Wang, Y., He, X., Peng, S., Tan, D., Zhou, X.: Efficient loftr: Semi-dense local feature matching with sparse-like speed. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21666–21675 (2024)

    2024

  49. [49]

    $\pi^3$: Permutation-Equivariant Visual Geometry Learning

    Wang, Y., Zhou, J., Zhu, H., Chang, W., Zhou, Y., Li, Z., Chen, J., Pang, J., Shen, C., He, T.:π 3: Permutation-equivariant visual geometry learning. arXiv preprint arXiv:2507.13347 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  50. [50]

    Wang, Y., Lipson, L., Deng, J.: Sea-raft: Simple, efficient, accurate raft for optical flow. In: Eur. Conf. Comput. Vis. pp. 36–54. Springer (2024)

    2024

  51. [51]

    Wu, R., Cheng, X., Zhu, J., Liu, X., Zhang, M., Yan, S.: Uavd4l: A large-scale dataset for uav 6-dof localization. In: Int. Conf. 3D Vis. (2024)

    2024

  52. [52]

    arXiv preprint arXiv:2405.11936 (2024) 22, 29

    Xu, W., Yao, Y., Cao, J., Wei, Z., Liu, C., Wang, J., Peng, M.: Uav-visloc: A large- scale dataset for uav visual localization. arXiv preprint arXiv:2405.11936 (2024)

    work page arXiv 2024

  53. [53]

    In: IEEE Conf

    Yan, Q., Zheng, J., Reding, S., Li, S., Doytchinov, I.: Crossloc: Scalable aerial localization assisted by multimodal synthetic data. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17358–17368 (2022)

    2022

  54. [54]

    Yan, S., Cheng, X., Liu, Y., Zhu, J., Wu, R., Liu, Y., Zhang, M.: Render-and- compare: Cross-view 6-dof localization from noisy prior. In: Int. Conf. Multimedia and Expo. pp. 2171–2176. IEEE (2023)

    2023

  55. [55]

    Advances in Neural Information Processing Systems37, 21875–21911 (2024)

    Yang, L., Kang, B., Huang, Z., Zhao, Z., Xu, X., Feng, J., Zhao, H.: Depth anything v2. Advances in Neural Information Processing Systems37, 21875–21911 (2024)

    2024

  56. [56]

    Exploring the best way for UAV visual localization under Low-altitude Multi-view Observation Condition: a Benchmark

    Ye, Y., Teng, X., Chen, S., Li, Z., Liu, L., Yu, Q., Tan, T.: Exploring the best way for uav visual localization under low-altitude multi-view observation condition: a benchmark. arXiv preprint arXiv:2503.10692 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  57. [57]

    Ufm: A simple path towards unified dense correspondence with flow.arXiv preprint arXiv:2506.09278, 2025

    Zhang, Y., Keetha, N., Lyu, C., Jhamb, B., Chen, Y., Qiu, Y., Karhade, J., Jha, S., Hu, Y., Ramanan, D., et al.: Ufm: A simple path towards unified dense corre- spondence with flow. arXiv preprint arXiv:2506.09278 (2025)

    work page arXiv 2025

  58. [58]

    In: IEEE/RSJ Int

    Zhang, Z., Gupta, A., Jiang, H., Singh, H.: Neuflow-v2: Push high-efficiency optical flow to the limit. In: IEEE/RSJ Int. Conf. Intell. Robots Syst. pp. 2479–2485. IEEE (2025)

    2025

  59. [59]

    In: ACM Int

    Zheng, Z., Wei, Y., Yang, Y.: University-1652: A multi-view multi-source bench- mark for drone-based geo-localization. In: ACM Int. Conf. Multimedia. pp. 1395– 1403 (2020) OrthoTrack 19

    2020

  60. [60]

    Zhu, J., Peng, S., Wang, L., Tan, H., Liu, Y., Zhang, M., Yan, S.: Lod-loc v2: Aerial visual localization over low level-of-detail city models using explicit silhou- ette alignment. In: Int. Conf. Comput. Vis. pp. 26610–26621 (2025)

    2025

  61. [61]

    In: Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C

    Zhu, J., Yan, S., Wang, L., shengYue, z., Liu, Y., Zhang, M.: Lod-loc: Aerial visual localization using lod 3d map with neural wireframe alignment. In: Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C. (eds.) Adv. Neural Inform. Process. Syst. vol. 37, pp. 119063–119098. Curran Associates, Inc. (2024) 20 O. Dhaouadi et a...

    2024