Research map / Label-efficient and robust learning / Recognition and long-tail
OrthoLoC: UAV 6-DoF Localization and Calibration Using Orthographic Geodata
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.
Introduces OrthoLoC, a 16.4K-image UAV dataset paired with orthophotos and elevation rasters for 6-DoF localization and calibration, benchmarks feature matchers, and proposes AdHoP, a homography-based refinement that improves any matcher.
Problem
Visual 6-DoF localization and calibration of UAV images using lightweight orthographic geodata (orthophotos and DSMs) instead of image databases or heavy 3D models, where no paired benchmark exists and perspective-to-orthographic domain gaps hinder matching.
Approach
Pair each UAV query with cropped orthophoto and DSM references derived from photogrammetric reconstructions to decouple retrieval from pose estimation, lift 2D-2D matches to 3D via DSM, and refine with an ad hoc homography (AdHoP) warp of the orthophoto that is method-agnostic.
Key results
- GIM+DKM with AdHoP achieves 75.4% recall at 1m-1° and 88.4% at 5m-5° on the OrthoLoC test set, the best among tested matchers.
- AdHoP reduces DKM translation error by 63% (3.83 m to 1.40 m) and improves XFeat* matching error by about 95% (222.2 px to 9.2 px).
- GIM+DKM with AdHoP achieves 1.6% relative focal length error with 2.09 m translation error in calibration, though recall is only 21.8%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add AdHoP refinement to GIM+DKM (ablation) | GIM+DKM · OrthoLoC test sets | translation error (m): 0.40 -> 0.32 | Helped |
| Add AdHoP refinement to XFeat* (ablation) | XFeat* · OrthoLoC test sets | matching error (px): 222.2 -> 9.2 | Helped |
| Add AdHoP refinement to DKM (ablation) | DKM · OrthoLoC test sets | translation error (m): 3.83 -> 1.40 | Helped |
| Add AdHoP refinement to RoMa (ablation) | RoMa · OrthoLoC test sets | 1m-1° recall (%): 44.4 -> 54.6 | Helped |
| Use cross-domain DOP instead of same-domain DOP | GIM+DKM+AdHoP · OrthoLoC cross-domain DOP | translation error | Hurt |
| Use cross-domain DOP and DSM instead of same-domain reference data | GIM+DKM+AdHoP · OrthoLoC cross-domain DOP and DSM | translation error (m): 0.16 -> 1.12 | Hurt |
| Scale images to 30% of original size | OrthoLoC localization methods · OrthoLoC resolution analysis | localization accuracy: 30% | No clear effect |
| Reduce covisibility below 20% | OrthoLoC localization methods · OrthoLoC covisibility analysis | localization accuracy: 20% | Hurt |
Surprising findings
- Semi-dense matchers LoFTR and XoFTR perform poorly on orthographic geodata, with recall below about 20%, despite strong general benchmark performance.
- Camera calibration suffers from geometric ambiguity between focal length and translation, limiting accuracy even with the best method.
Limitations
- Calibration is affected by focal length and translation ambiguity, and recall remains low (21.8%) for calibration.
- Requires geodata with at least 20% covisibility and does not scale to large rasters without region detection.
- AdHoP fails when matches are completely corrupted, since it only improves partially incorrect correspondences.
Experimental setup
Datasets: OrthoLoC AnyVisLoc
Benchmarks: OrthoLoC
Metrics: median error (ME) px translation error (TE) m rotation error (RE) deg reprojection error (RPE) px recall at 1m-1°, 3m-3°, 5m-5° relative focal error (RFE) % speed s
Baselines: SuperPoint+SuperGlue SuperPoint+LightGlue DeDoDe XFeat LoFTR eLoFTR XoFTR DKM DUSt3R MASt3R RoMa
Builds on: LoFTR GIM DKM RoMa MASt3R
Closest papers by meaning
- Game4Loc: A UAV Geo-Localization Benchmark from Game Data (AAAI 2025 · similarity 0.88)
- Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery (NeurIPS 2026 · similarity 0.85)
- View From Above: Orthogonal-View aware Cross-view Localization (CVPR 2024 · similarity 0.85)
- SOAR: Regression-based LiDAR Relocalization for UAVs (NeurIPS 2026 · similarity 0.85)
- Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images (NeurIPS 2026 · similarity 0.85)
- LoD-Loc: Aerial Visual Localization using LoD 3D Map with Neural Wireframe Alignment (NeurIPS 2024 · 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.