Research map / 3D vision / Depth and stereo
LoD-Loc: Aerial Visual Localization using LoD 3D Map with Neural Wireframe Alignment
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 LoD-Loc, which localizes UAV images against LoD city models by aligning projected building wireframes with network-predicted wireframe probability maps via coarse-to-fine cost-volume pose selection and differentiable Gauss-Newton refinement.
Problem
Aerial 6-DoF visual localization of UAV cameras using lightweight, privacy-preserving LoD 3D city models, which lack texture and make point-feature 2D-3D matching nearly impossible.
Approach
Replace texture-based 2D-3D correspondence with wireframe alignment: a U-Net predicts a single-channel wireframe probability map, a hierarchical 4-DoF pose cost volume (with uncertainty-based adaptive sampling range) selects the pose, and a differentiable Gauss-Newton step refines all 6 DoF, trainable end-to-end with pose supervision.
Key results
- On UAVD4L-LoD in-Traj. at 5m-5°, full LoD-Loc reaches 96.95% recall, versus 93.77% for UAVD4L with RoMA (texture mesh) and 3.68% for CadLoc with RoMA.
- On Swiss-EPFL in-Place at 2m-2°, LoD-Loc achieves 48.60% recall, above UAVD4L with RoMA at 45.98%, but below RoMA on out-of-Place queries (37.73% vs 89.18% at 2m-2°).
- Pose refinement improves median error on UAVD4L-LoD in-Traj. from 1.97 m / 1.25° at level 3 to 0.97 m / 0.52°; the method runs in 0.34 s per query versus 1.7–5.4 s for baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use neural wireframe estimation instead of DeepLSD line extraction and distance fields (ablation) | LoD-Loc · UAVD4L-LoD in-Traj. | recall (2m-2°): 10.41 -> 84.41 | Helped |
| Add uncertainty sampling range estimation (ablation) | LoD-Loc · UAVD4L-LoD in-Traj. | recall (2m-2°): 70.39 -> 84.41 | Helped |
| Add pose refinement (ablation) | LoD-Loc · UAVD4L-LoD in-Traj. | recall (2m-2°): 51.31 -> 84.41 | Helped |
| Increase pose sampling numbers from [2, 3, 3, 8] to [8, 10, 10, 30] (ablation) | LoD-Loc · UAVD4L-LoD in-Traj. | recall (2m-2°): 18.83 -> 84.41 | Helped |
| Increase pose sampling numbers from [2, 3, 3, 8] to [8, 10, 10, 30] (ablation) | LoD-Loc · UAVD4L-LoD out-of-Traj. | recall (2m-2°): 12.36 -> 95.94 | Helped |
Surprising findings
- Off-the-shelf line-based and point-based matchers on LoD models (CadLoc with SOLD2, DeepLSD+GlueStick, LoFTR, SIFT) almost completely fail, with near-zero recall.
- On Swiss-EPFL out-of-Place queries, LoD-Loc falls well behind the texture-mesh UAVD4L baseline with RoMA.
Limitations
- Assumes a known gravity direction and a location prior from sensors, so it does not work where GPS is denied.
- Performance on the LoD2.0 Swiss-EPFL dataset is weaker, especially out-of-Place, since LoD2.0 has less structural detail.
- Accuracy degrades sharply when the initial pose error grows (e.g., recall falls to 5.67% at a [100,100,30] m prior error range).
Experimental setup
Datasets: UAVD4L-LoD Swiss-EPFL UAVD4L CrossLoc LHS
Benchmarks: UAVD4L-LoD Swiss-EPFL
Metrics: recall at 2m-2° recall at 3m-3° recall at 5m-5° median translation error median rotation error
Baselines: UAVD4L CadLoc SIFT+NN SuperPoint+SuperGlue LoFTR e-LoFTR RoMA SOLD2 DeepLSD+GlueStick
Builds on: U-Net DeepLSD Gauss-Newton
Closest papers by meaning
- Game4Loc: A UAV Geo-Localization Benchmark from Game Data (AAAI 2025 · similarity 0.85)
- Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images (NeurIPS 2026 · similarity 0.84)
- OrthoLoC: UAV 6-DoF Localization and Calibration Using Orthographic Geodata (NeurIPS 2025 · similarity 0.84)
- FG^2: Fine-Grained Cross-View Localization by Fine-Grained Feature Matching (CVPR 2025 · similarity 0.84)
- UA-Pose: Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References (CVPR 2025 · similarity 0.84)
- Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery (NeurIPS 2026 · similarity 0.84)
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.