Research map / Point clouds / Point cloud classification
Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes "ray cloud," a geometric lifting of 3D points to rays through two centers, enabling real-time privacy-preserving visual localization via a 5+1-point relative pose solver, plus an adapted inversion attack and voxel-based ray sampling defense.
Problem
Privacy-preserving visual localization: sparse SfM point clouds can be inverted into scene images, and existing line-cloud defenses are too slow for real-time single-image localization due to weak geometric constraints.
Approach
Lift each 3D point to a ray passing through one of two fixed centers (K-means centers), so localization becomes relative pose estimation against a calibrated two-camera rig (p5+1R solver, 5-point algorithm plus one extra correspondence to fix scale), and defend against density-based inversion with a modified attack and voxel-based checkerboard ray assignment.
Key results
- Ray cloud localization runs at about 137 FPS on average (voxel-based ray cloud 136 FPS), versus roughly 3-11 FPS for PPL and uniform line cloud baselines, with similar pose accuracy (e.g., Energy Landscape mean rotation error 0.262 deg vs 0.292 for PPL).
- On Energy Landscape the ray cloud yields PSNR 10.29 and SSIM 0.319 for recovered images, versus 10.42 and 0.355 for PPL, indicating degraded inversion quality.
- On Cambridge, ray cloud has larger pose errors (mean translation 0.219 m vs 0.181 m for PPL) but a 239 ms... runtime figure of 12 ms median runtime-per-query is reported in Table 1 versus 186 ms for PPL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the ray-cloud representation and p5+1R solver instead of PPL | Visual localization · Energy Landscape | runtime (ms): 88 -> 6 | Helped |
| Use the ray-cloud representation and p5+1R solver instead of PPL | Visual localization · Cambridge | runtime (ms): 186 -> 12 | Helped |
| Use ray cloud instead of PPL | Visual localization · Energy Landscape | rotation error (°): 0.292 -> 0.262 | Helped |
| Use ray cloud instead of PPL | Visual localization · Cambridge | rotation error (°): 0.775 -> 0.855 | Hurt |
| Add voxel-based ray sampling (ablation) | Ray cloud · Energy Landscape | geometric error (m): 10.29 -> 10.05 | Helped |
| Add ray rejection sampling at β=25% (ablation) | Ray cloud inversion attack · Energy Landscape | geometric error (m): 1.027 -> 0.669 | Helped |
| Add ray rejection sampling at β=25% (ablation) | Ray cloud inversion attack · Cambridge | geometric error (m): 11.47 -> 9.614 | Helped |
| Set ray rejection threshold to 50% instead of 25% (ablation) | Ray cloud inversion attack · Cambridge | geometric error (m): 9.614 -> 12.98 | Hurt |
Surprising findings
- The ray cloud has slightly higher localization error than uniform line cloud and PPL on the outdoor Cambridge dataset, attributed to the larger baseline between ray centers.
- Voxel-based ray sampling did not significantly reduce pose error compared with the standard ray cloud and can degrade scene recovery quality only in some cases.
Limitations
- Potentially vulnerable to an attack that unfolds projected rays onto cylindrical or spherical image planes and feeds them to an inversion network.
- Rejection sampling adds a hyperparameter (beta) that must be tuned, and voxel size must be manually chosen per dataset.
- Higher localization error than baselines on outdoor Cambridge scenes.
Experimental setup
Datasets: Energy Landscape Cambridge Landmarks
Benchmarks: Energy Landscape Cambridge Landmarks
Metrics: rotation error translation error runtime inlier ratio median point error PSNR SSIM MAE
Baselines: Point cloud (p3P) Uniform line cloud (p6L) PPL PPL+ Pan et al. permutation lifting
Builds on: uniform line cloud Paired-Point Lifting (PPL) geometry inversion attack (Chelani et al.) InvSfM LO-RANSAC
Closest papers by meaning
- Gaussian Splatting Feature Fields for (Privacy-Preserving) Visual Localization (CVPR 2025 · similarity 0.86)
- ConcreTizer: Model Inversion Attack via Occupancy Classification and Dispersion Control for 3D Point Cloud Restoration (ICLR 2025 · similarity 0.84)
- GS-Hider: Hiding Messages into 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.83)
- SecureGS: Boosting the Security and Fidelity of 3D Gaussian Splatting Steganography (ICLR 2025 · similarity 0.83)
- Explicitly Perceiving and Preserving the Local Geometric Structures for 3D Point Cloud Attack (AAAI 2024 · similarity 0.82)
- 360Loc: A Dataset and Benchmark for Omnidirectional Visual Localization with Cross-device Queries (CVPR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.