pith. sign in

arxiv: 2604.22310 · v1 · submitted 2026-04-24 · 💻 cs.CV

Revisiting Geometric Obfuscation with Dual Convergent Lines for Privacy-Preserving Image Queries in Visual Localization

Jeonggon Kim , Heejoon Moon , Je Hyeong Hong This is my paper

Pith reviewed 2026-05-08 12:47 UTC · model grok-4.3

classification 💻 cs.CV
keywords privacy-preserving image queriesvisual localizationkeypoint obfuscationgeometric obfuscationdual convergent linespose estimationcloud-based localizationattack resilience
0
0 comments X

The pith

Dual Convergent Lines route each keypoint to a line from one of two fixed anchors on a partition line, rendering geometry-recovery attacks ill-posed.

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

The paper seeks to fix a core weakness in geometry-based privacy protection for cloud visual localization. Previous obfuscation methods leave neighboring lines still surrounding the true keypoint, so attacks can recover the original point by solving for a consistent intersection. Dual Convergent Lines instead fixes two anchors on a central partition and sends each keypoint along a line from whichever anchor lies on its side. Lines from the same side converge to the wrong anchor and give a trivial false solution; lines that cross the boundary become nearly parallel and give wildly varying solutions. This change works with standard line solvers and keeps pose estimation accurate on indoor and large outdoor datasets.

Core claim

DCL places two fixed anchors on a central partition line and lifts each keypoint to a line originating from one of them, with the active anchor determined by the keypoint's location. Neighboring lines either misleadingly converge to one anchor, yielding a trivial solution, or become near-parallel at the partition boundary, yielding an unstable high-variance solution. Both outcomes thwart point recovery while remaining compatible with existing line-based localization solvers.

What carries the argument

Dual Convergent Lines (DCL), the mechanism that assigns each obfuscated line to one of two fixed anchors on a partition line according to the keypoint's side.

If this is right

  • Standard line-based pose solvers can ingest DCL features directly without code changes.
  • The same obfuscation works on both small indoor scenes and large-scale outdoor maps.
  • No additional segmentation step is required to achieve attack resistance.
  • Localization pipelines gain a drop-in privacy layer that preserves metric accuracy.

Where Pith is reading between the lines

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

  • The partition idea could be tested on 3D point clouds or other geometric features where surrounding distributions enable recovery.
  • Adaptive anchor placement might be needed if future attacks specifically target the boundary region.
  • Disrupting convex-hull coverage of lines may be a reusable principle for other geometric privacy tasks.

Load-bearing premise

Prior geometry-based attacks rely on the spatial distribution of obfuscated neighboring lines surrounding the original keypoint, and the dual-anchor partition disrupts recovery without introducing new exploitable patterns or substantially harming localization performance.

What would settle it

An experiment in which a geometry-recovery attack still locates the original keypoints with low error on DCL-obfuscated data, or in which localization accuracy drops sharply compared with non-obfuscated keypoints.

Figures

Figures reproduced from arXiv: 2604.22310 by Heejoon Moon, Je Hyeong Hong, Jeonggon Kim.

Figure 1
Figure 1. Figure 1: Existing privacy-preserving schemes such as Random view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of our DCL construction procedure. We begin by extracting keypoints [ view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of DCL’s two failure modes against the view at source ↗
Figure 5
Figure 5. Figure 5: Results of revealed Superpoint [12] keypoints from the geometry-recovery attck [7]. In (d), the view is zoomed out by ×2 to illustrate that the recovered points from DCL largely fall outside the original image region. opposite region, R2, and only these neighboring lines avoid trivially converging at the anchor a1. However, as keypoints from R1 and R2 approach the central boundary, their cor￾responding DCL… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative inversion results with the Superpoint [ view at source ↗
Figure 7
Figure 7. Figure 7: Example of server-side attack, where inlier feature po view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of the iterative server-side attack on three view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative inversion results using SuperPoint features against the geometry-recovery attack [7] in an expanded neighborhood (K = 100) setting. The reconstructed images are organized by dataset: Row 1 (Aachen), Rows 2–3 (Cambridge), and Rows 4–5 (7Scenes). Feature positions are recovered via various methods: (a) Feature Points, (b) Random Lines [51], (c) Coordinate Permutation [34], and (d) Dual Convergent… view at source ↗
Figure 10
Figure 10. Figure 10: Visualization of the inversion network’s output using view at source ↗
Figure 11
Figure 11. Figure 11: Inversion results generated by the InvSfM [40] model using SIFT features, conducted in the oracle setting with K = 20. The reconstructed images are organized by dataset: Row 1 (Aachen), Rows 2–3 (Cambridge), and Rows 4–5 (7Scenes). Feature positions are recovered via various methods: (a) Feature Points, (b) Random Lines [51], (c) Coordinate Permutation [34], and (d) Dual Convergent Lines (ours) view at source ↗
read the original abstract

Privacy-Preserving Image Queries (PPIQ) are an emerging mechanism for cloud-based visual localization, enabling pose estimation from obfuscated features instead of private images or raw keypoints. However, the main approaches for PPIQ, primarily geometry-based and segmentation-based obfuscation, both suffer from vulnerabilities to recent privacy attacks. In particular, a fundamental limitation of geometry-based obfuscation is that the spatial distribution of obfuscated neighboring lines still effectively surrounds the original keypoint location, providing exploitable cues for recovering the original points. We revisit this geometric paradigm and introduce Dual Convergent Lines (DCL), a novel keypoint obfuscation method demonstrating strong resilience against such attack. DCL places two fixed anchors on a central partition line and lifts each keypoint to a line originating from one of them, with the active anchor determined by the keypoint's location. This arrangement invalidates the geometry-recovery attack by making its optimization ill-posed: Neighboring lines either misleadingly converge to one anchor, yielding a trivial solution, or become near-parallel at the partition boundary, yielding an unstable high-variance solution. Both outcomes thwart point recovery. DCL is also compatible with an existing line-based solver, enabling deployment in traditional localization pipelines. Experiments on both indoor and large-scale outdoor datasets demonstrate DCL's robustness against privacy attacks, efficiency, and scalability, while achieving practical localization performance.

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

2 major / 2 minor

Summary. The paper proposes Dual Convergent Lines (DCL) as a geometric obfuscation technique for privacy-preserving image queries in visual localization. It places two fixed anchors on a central partition line and lifts each keypoint to a line from one of the anchors (chosen by keypoint location), claiming this renders single-origin geometry-recovery attacks ill-posed (trivial convergence or high-variance near-parallel solutions) while remaining compatible with line-based solvers and delivering practical localization accuracy on indoor and large-scale outdoor datasets.

Significance. If the core resilience claim holds under adapted attacks, DCL would strengthen the geometric obfuscation paradigm for PPIQ by directly addressing the spatial-distribution vulnerability identified in prior work. The construction is novel and the compatibility with existing solvers is a practical strength. However, the absence of analysis or experiments against modified optimizers (e.g., joint anchor selection) and the lack of error bars or raw data in the reported experiments limit the assessed significance; the result is promising but currently rests on an untested assumption about attack non-adaptability.

major comments (2)
  1. [Abstract / attack analysis] Abstract and attack-resilience section: the central claim that DCL 'invalidates the geometry-recovery attack by making its optimization ill-posed' assumes the attacker uses the unmodified single-origin formulation. No formal argument, ablation, or experiment is provided against an adapted optimizer that introduces a per-line anchor-choice variable or exploits the known central partition line as a disambiguation cue. This directly undermines the 'thwarts point recovery' conclusion and is load-bearing for the paper's main contribution.
  2. [Experiments] Experiments section: robustness is asserted on indoor and outdoor datasets, yet no quantitative comparison to prior geometric obfuscation baselines, no error bars on localization or attack-success metrics, and no raw data or full derivation details are referenced. This makes it impossible to verify that the dual-anchor arrangement does not introduce new exploitable patterns or degrade performance beyond acceptable limits.
minor comments (2)
  1. [Method] The free parameters (anchor positions and partition-line placement) are listed but their selection procedure and sensitivity analysis are not detailed; a short paragraph or table would clarify reproducibility.
  2. [Method] Notation for the partition line and anchor selection rule should be formalized with an equation or pseudocode to avoid ambiguity when describing the 'active anchor determined by the keypoint's location'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment below with clarifications on the scope of our claims and outline specific revisions to improve the presentation of attack analysis and experimental details.

read point-by-point responses

  1. Referee: [Abstract / attack analysis] Abstract and attack-resilience section: the central claim that DCL 'invalidates the geometry-recovery attack by making its optimization ill-posed' assumes the attacker uses the unmodified single-origin formulation. No formal argument, ablation, or experiment is provided against an adapted optimizer that introduces a per-line anchor-choice variable or exploits the known central partition line as a disambiguation cue. This directly undermines the 'thwarts point recovery' conclusion and is load-bearing for the paper's main contribution.

    Authors: We thank the referee for this observation. The manuscript's core claim targets the single-origin geometry-recovery attack from prior work, where the dual-anchor construction forces neighboring lines to either converge trivially to one fixed anchor or produce near-parallel unstable solutions near the partition boundary. This directly addresses the spatial-distribution vulnerability noted in earlier geometric obfuscation methods. We did not include experiments or formal analysis against an adapted optimizer that jointly optimizes anchor selection or exploits the known partition line, as our evaluation focused on resilience to the established attack formulation. We agree this is a valuable point for strengthening the contribution. In the revised manuscript, we will expand the attack-resilience section with a discussion of why consistent anchor choice across lines remains challenging even under adaptation (due to the fixed anchors and location-based selection rule) and add an ablation on simulated adapted attacks where feasible. revision: yes

  2. Referee: [Experiments] Experiments section: robustness is asserted on indoor and outdoor datasets, yet no quantitative comparison to prior geometric obfuscation baselines, no error bars on localization or attack-success metrics, and no raw data or full derivation details are referenced. This makes it impossible to verify that the dual-anchor arrangement does not introduce new exploitable patterns or degrade performance beyond acceptable limits.

    Authors: We appreciate the referee's call for greater experimental transparency. The current experiments demonstrate DCL's localization accuracy and attack resilience on indoor and large-scale outdoor datasets, showing compatibility with line-based solvers and practical performance. To address the gaps, the revised version will add direct quantitative comparisons against prior geometric obfuscation baselines (e.g., single convergent line methods), include error bars on all localization error and attack-success metrics, and reference or include raw data and derivation details in the supplementary material. These changes will enable better verification that the dual-anchor design does not introduce new patterns while maintaining acceptable performance. revision: yes

Circularity Check

0 steps flagged

No significant circularity; DCL is a direct geometric construction

full rationale

The paper defines Dual Convergent Lines explicitly via two fixed anchors on a partition line and assigns each keypoint to one anchor based on location. It then argues that this specific arrangement renders the prior single-origin attack optimization ill-posed (trivial convergence or high-variance near-parallel lines). This follows directly from the construction and the attack's stated assumptions rather than any self-referential fit, renamed prediction, or load-bearing self-citation chain. No equations reduce the claimed resilience to the input definition by construction, and the method is presented as compatible with existing solvers without circular justification. The central claim therefore retains independent geometric content.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The central claim rests on domain assumptions about attack geometry and introduces a new obfuscation construction whose parameters and effectiveness are validated only through the paper's own experiments.

free parameters (2)
  • Anchor positions on partition line
    Two fixed anchors whose exact locations must be chosen or tuned to achieve the convergent/parallel behavior.
  • Central partition line placement and orientation
    Determines side-based anchor selection and boundary behavior for each keypoint.
axioms (1)
  • domain assumption Geometry-recovery attacks exploit surrounding spatial distribution of obfuscated neighboring lines around original keypoints
    Invoked to explain the core limitation of prior geometric obfuscation methods.
invented entities (1)
  • Dual Convergent Lines (DCL) obfuscation no independent evidence
    purpose: To render geometry-based point recovery ill-posed while preserving localization utility
    New geometric construction introduced to address the stated attack vulnerability.

pith-pipeline@v0.9.0 · 5555 in / 1444 out tokens · 48519 ms · 2026-05-08T12:47:54.233427+00:00 · methodology

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

57 extracted references · 57 canonical work pages

  1. [1]

    Netvlad: Cnn architecture for weakly supervised place recognition

    Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pa- jdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5297–5307, 2016. 7

    work page 2016

  2. [2]

    Visual camera re- localization from RGB and RGB-D images using DSAC

    Eric Brachmann and Carsten Rother. Visual camera re- localization from RGB and RGB-D images using DSAC. IEEE Transactions on Pattern Analysis and Machine Intel- ligence (TPAMI), 2021. 7, 13

    work page 2021

  3. [3]

    Dsac-differentiable ransac for camera localization

    Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6684–6692,

    work page

  4. [4]

    On the limits of pseudo ground truth in vi- sual camera re-localisation

    Eric Brachmann, Martin Humenberger, Carsten Rother, and Torsten Sattler. On the limits of pseudo ground truth in vi- sual camera re-localisation. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6218–6228, 2021. 14

    work page 2021

  5. [5]

    Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses

    Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5044–5053, 2023. 7, 8, 13, 14

    work page 2023

  6. [6]

    Privacy-preserving representations are not enough: Recovering scene content from camera poses

    Kunal Chelani, Torsten Sattler, Fredrik Kahl, and Zuzana Kukelova. Privacy-preserving representations are not enough: Recovering scene content from camera poses. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 13132–13141,

    work page

  7. [7]

    Obfuscation based privacy preserving representations are recoverable using neighborhood informa- tion

    Kunal Chelani, Assia Benbihi, Fredrik Kahl, Torsten Sattler, and Zuzana Kukelova. Obfuscation based privacy preserving representations are recoverable using neighborhood informa- tion. InInternational Conference on 3D Vision (3DV), 2025. 1, 2, 3, 4, 5, 6, 7, 8, 13, 14, 15, 16, 17

    work page 2025

  8. [8]

    Locally opti- mized ransac

    Ondrej Chum, Jir ´ı Matas, and Josef Kittler. Locally opti- mized ransac. InPattern Recognition: 25th DAGM Sympo- sium, Magdeburg, Germany, September 10-12, 2003. Pro- ceedings 25, pages 236–243. Springer, 2003. 7

    work page 2003

  9. [9]

    Lee, Hyo Jin Kim, Tianwei Shen, Meghan Cowan, Rajvi Shah, Caroline Trippel, Bran- don Reagen, Timothy Sherwood, Vasileios Balntas, Armin Alaghi, and Eddy Ilg

    Deeksha Dangwal, Vincent T. Lee, Hyo Jin Kim, Tianwei Shen, Meghan Cowan, Rajvi Shah, Caroline Trippel, Bran- don Reagen, Timothy Sherwood, Vasileios Balntas, Armin Alaghi, and Eddy Ilg. Analysis and mitigations of reverse engineering attacks on local feature descriptors, 2021. 13

    work page 2021

  10. [10]

    Mitigating reverse engineering attacks on local feature de- scriptors

    Deeksha Dangwal, Vincent T Lee, Hyo Jin Kim, Tianwei Shen, Meghan Cowan, Rajvi Shah, Caroline Trippel, Bran- don Reagen, Timothy Sherwood, Vasileios Balntas, et al. Mitigating reverse engineering attacks on local feature de- scriptors. InBritish Machine Vision Conference (BMVC),

    work page

  11. [11]

    Davison, Ian D

    Andrew J. Davison, Ian D. Reid, Nicholas D. Molton, and Olivier Stasse. MonoSLAM: Real-time single camera SLAM.IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 29(6):1052–1067, 2007. 1

    work page 2007

  12. [12]

    Superpoint: Self-supervised interest point detection and description

    Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabi- novich. Superpoint: Self-supervised interest point detection and description. InProceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition workshops (CVPRW), pages 224–236, 2018. 2, 3, 5, 6, 7, 8, 13

    work page 2018

  13. [13]

    Privacy-preserving image features via adversarial affine subspace embeddings

    Mihai Dusmanu, Johannes L Schonberger, Sudipta N Sinha, and Marc Pollefeys. Privacy-preserving image features via adversarial affine subspace embeddings. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14267–14277, 2021. 1, 2

    work page 2021

  14. [14]

    Fischler and Robert C

    Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography.Communications of the ACM, 24(6):381–395, 1981. 3

    work page 1981

  15. [15]

    Privacy preserving local- ization and mapping from uncalibrated cameras

    Marcel Geppert, Viktor Larsson, Pablo Speciale, Johannes L Schonberger, and Marc Pollefeys. Privacy preserving local- ization and mapping from uncalibrated cameras. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1809–1819, 2021. 2

    work page 2021

  16. [16]

    Privacy preserving partial localiza- tion

    Marcel Geppert, Viktor Larsson, Johannes L Sch ¨onberger, and Marc Pollefeys. Privacy preserving partial localiza- tion. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 17337– 17347, 2022. 2

    work page 2022

  17. [17]

    Using Global Localization to Improve Navigation

    Google. Using Global Localization to Improve Navigation. https://ai.googleblog.com/2019/02/using- global-localization-to-improve.html, 2019. 1

    work page 2019

  18. [18]

    Scope of va- lidity of psnr in image/video quality assessment.Electronics letters, 44(13):800–801, 2008

    Quan Huynh-Thu and Mohammed Ghanbari. Scope of va- lidity of psnr in image/video quality assessment.Electronics letters, 44(13):800–801, 2008. 6

    work page 2008

  19. [19]

    PoseNet: A convolutional network for real-time 6-DOF cam- era relocalization

    Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF cam- era relocalization. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), pages 2938– 2946, 2015. 2, 6, 7, 14

    work page 2015

  20. [20]

    Re- visiting Geometric Obfuscation for Privacy-Preserving Im- age Queries in Visual Localization

    Jeonggon Kim, Heejoon Moon, and Je Hyeong Hong. Re- visiting Geometric Obfuscation for Privacy-Preserving Im- age Queries in Visual Localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026. 14, 15

    work page 2026

  21. [21]

    Effi- cient intersection of three quadrics and applications in com- puter vision

    Zuzana Kukelova, Jan Heller, and Andrew Fitzgibbon. Effi- cient intersection of three quadrics and applications in com- puter vision. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1799–1808, 2016. 3

    work page 2016

  22. [22]

    PoseLib - Minimal Solvers for Camera Pose Estimation.https://github.com/vlarsson/ PoseLib, 2020

    Viktor Larsson. PoseLib - Minimal Solvers for Camera Pose Estimation.https://github.com/vlarsson/ PoseLib, 2020. Accessed: 2022-10-30. 7, 14

    work page 2020

  23. [23]

    Paired-point lifting for enhanced privacy- preserving visual localization

    Chunghwan Lee, Jaihoon Kim, Chanhyuk Yun, and Je Hyeong Hong. Paired-point lifting for enhanced privacy- preserving visual localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17266–17275, 2023. 2

    work page 2023

  24. [24]

    A method for the solution of certain non-linear problems in least squares.Quarterly of applied mathematics, 2(2):164–168, 1944

    Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares.Quarterly of applied mathematics, 2(2):164–168, 1944. 7

    work page 1944

  25. [25]

    David G. Lowe. Distinctive image features from scale- invariant keypoints.International Journal of Computer Vi- sion, 60(2):91–110, 2004. 2, 3, 6, 7, 13, 16, 17, 18

    work page 2004

  26. [26]

    Get out of my lab: Large-scale, real-time visual-inertial localization

    Simon Lynen, Torsten Sattler, Michael Bosse, Joel A Hesch, Marc Pollefeys, and Roland Siegwart. Get out of my lab: Large-scale, real-time visual-inertial localization. In Robotics: Science and Systems, page 1, 2015. 1

    work page 2015

  27. [27]

    Azure Spatial Anchors.https://azure

    Microsoft. Azure Spatial Anchors.https://azure. microsoft . com / en - us / products / spatial - anchors, 2024. 1

    work page 2024

  28. [28]

    Ef- ficient privacy-preserving visual localization using 3d ray clouds

    Heejoon Moon, Chunghwan Lee, and Je Hyeong Hong. Ef- ficient privacy-preserving visual localization using 3d ray clouds. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9773–9783, 2024. 2, 3

    work page 2024

  29. [29]

    Depth-guided privacy-preserving visual localization using 3d sphere clouds

    Heejoon Moon, Jongwoo Lee, Jeonggon Kim, and Je Hyeong Hong. Depth-guided privacy-preserving visual localization using 3d sphere clouds. InBritish Machine Vi- sion Conference (BMVC), 2024. 2

    work page 2024

  30. [30]

    ORB-SLAM: A versatile and accurate monocular SLAM system.IEEE Transactions on Robotics, 31(5):1147– 1163, 2015

    Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. ORB-SLAM: A versatile and accurate monocular SLAM system.IEEE Transactions on Robotics, 31(5):1147– 1163, 2015. 1

    work page 2015

  31. [31]

    NinjaDesc: Content-concealing visual descriptors via adversarial learn- ing

    Tony Ng, Hyo Jin Kim, Vincent T Lee, Daniel DeTone, Tsun-Yi Yang, Tianwei Shen, Eddy Ilg, Vassileios Bal- ntas, Krystian Mikolajczyk, and Chris Sweeney. NinjaDesc: Content-concealing visual descriptors via adversarial learn- ing. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 12797– 12807, 2022. 1, 2, 13

    work page 2022

  32. [32]

    Niantic research presents: State of the art in relocalization with machine learning, 2023

    Niantic Labs. Niantic research presents: State of the art in relocalization with machine learning, 2023. 14

    work page 2023

  33. [33]

    Niantic spatial vps.https://www

    Niantic Spatial, Inc. Niantic spatial vps.https://www. nianticspatial . com / docs / nsdk / features / lightship_vps/, 2025. 1, 14

    work page 2025

  34. [34]

    Sch ¨onberger, Viktor Larsson, and Marc Pollefeys

    Linfei Pan, Johannes L. Sch ¨onberger, Viktor Larsson, and Marc Pollefeys. Privacy preserving localization via coor- dinate permutations. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV), pages 18174–18183, 2023. 1, 2, 5, 6, 7, 13, 14, 16, 17, 18

    work page 2023

  35. [35]

    Segloc: Learning segmentation-based representations for privacy-preserving visual localization

    Maxime Pietrantoni, Martin Humenberger, Torsten Sattler, and Gabriela Csurka. Segloc: Learning segmentation-based representations for privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 15380–15391,

    work page

  36. [36]

    Gaussian splatting feature fields for (privacy-preserving) vi- sual localization

    Maxime Pietrantoni, Gabriela Csurka, and Torsten Sattler. Gaussian splatting feature fields for (privacy-preserving) vi- sual localization. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 1082–1092, 2025. 1, 2, 7, 8, 14

    work page 2025

  37. [37]

    Vulnerability of privacy-preserving visual localization against diffusion-based attacks.https://openreview

    Maxime Pietrantoni, Torsten Sattler, and Gabriela Csurka. Vulnerability of privacy-preserving visual localization against diffusion-based attacks.https://openreview. net/forum?id=NmWf0gLufZ, 2025. 2, 3, 7, 8, 14

    work page 2025

  38. [38]

    Ldp-feat: Image features with local differential privacy

    Francesco Pittaluga and Bingbing Zhuang. Ldp-feat: Image features with local differential privacy. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17580–17590, 2023. 1, 2

    work page 2023

  39. [39]

    Learning privacy preserving encodings through adversarial training

    Francesco Pittaluga, Sanjeev Koppal, and Ayan Chakrabarti. Learning privacy preserving encodings through adversarial training. In2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 791–799. IEEE, 2019. 3

    work page 2019

  40. [40]

    Revealing scenes by inverting struc- ture from motion reconstructions

    Francesco Pittaluga, Sanjeev J Koppal, Sing Bing Kang, and Sudipta N Sinha. Revealing scenes by inverting struc- ture from motion reconstructions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 145–154, 2019. 1, 2, 3, 7, 8, 13, 14, 18

    work page 2019

  41. [41]

    A theory of minimal 3d point to 3d plane registration and its generaliza- tion.International journal of computer vision, 102(1):73–90,

    Srikumar Ramalingam and Yuichi Taguchi. A theory of minimal 3d point to 3d plane registration and its generaliza- tion.International journal of computer vision, 102(1):73–90,

    work page

  42. [42]

    P 2π: A minimal solution for registration of 3d points to 3d planes

    Srikumar Ramalingam, Yuichi Taguchi, Tim K Marks, and Oncel Tuzel. P 2π: A minimal solution for registration of 3d points to 3d planes. InProceedings of the European Confer- ence on Computer Vision (ECCV), pages 436–449. Springer,

    work page

  43. [43]

    U- Net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- Net: Convolutional networks for biomedical image segmen- tation. InInternational Conference on Medical Image Com- puting and Computer-Assisted Intervention (MICCAI), pages 234–241. Springer, 2015. 2

    work page 2015

  44. [44]

    From coarse to fine: Robust hierarchical localization at large scale

    Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12716–12725, 2019. 2, 6, 7, 8, 13, 14

    work page 2019

  45. [45]

    Superglue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4938–4947, 2020. 7, 8

    work page 2020

  46. [46]

    Image retrieval for image-based localization revis- ited

    Torsten Sattler, Tobias Weyand, Bastian Leibe, and Leif Kobbelt. Image retrieval for image-based localization revis- ited. InBritish Machine Vision Conference (BMVC), page 4,

    work page

  47. [47]

    Benchmark- ing 6dof outdoor visual localization in changing conditions

    Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmark- ing 6dof outdoor visual localization in changing conditions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8601–8610,

    work page

  48. [48]

    Structure-from-motion revisited

    Johannes Lutz Sch ¨onberger and Jan-Michael Frahm. Structure-from-motion revisited. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6

    work page 2016

  49. [49]

    Scene co- ordinate regression forests for camera relocalization in rgb- d images

    Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene co- ordinate regression forests for camera relocalization in rgb- d images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2930–2937, 2013. 2, 6, 8

    work page 2013

  50. [50]

    Privacy preserving image-based localization

    Pablo Speciale, Johannes L Schonberger, Sing Bing Kang, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image-based localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5493–5503, 2019. 2, 7

    work page 2019

  51. [51]

    Privacy preserving image queries for camera localization

    Pablo Speciale, Johannes L Schonberger, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image queries for camera localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1486–1496, 2019. 1, 2, 3, 5, 6, 7, 8, 13, 14, 16, 17, 18

    work page 2019

  52. [52]

    24/7 place recognition by view synthesis

    Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1808–1817, 2015. 7

    work page 2015

  53. [53]

    Glace: Global local accelerated coordinate encoding

    Fangjinhua Wang, Xudong Jiang, Silvano Galliani, Christoph V ogel, and Marc Pollefeys. Glace: Global local accelerated coordinate encoding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 7, 8, 13

    work page 2024

  54. [54]

    Dgc-gnn: leveraging geometry and color cues for visual descriptor-free 2d-3d matching

    Shuzhe Wang, Juho Kannala, and Daniel Barath. Dgc-gnn: leveraging geometry and color cues for visual descriptor-free 2d-3d matching. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 20881–20891, 2024. 2, 3, 7, 8

    work page 2024

  55. [55]

    Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Process- ing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Process- ing, 13(4):600–612, 2004. 6

    work page 2004

  56. [56]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586–595, 2018. 6

    work page 2018

  57. [57]

    Is geometry enough for matching in visual lo- calization? InProceedings of the European Conference on Computer Vision (ECCV), pages 407–425

    Qunjie Zhou, S ´ergio Agostinho, Aljosa Osep, and Laura Leal-Taix´e. Is geometry enough for matching in visual lo- calization? InProceedings of the European Conference on Computer Vision (ECCV), pages 407–425. Springer, 2022. 2, 3, 7, 8, 14 Revisiting Geometric Obfuscation with Dual Convergent Lines for Privacy-Preserving Image Queries in Visual Localizat...

    work page 2022