MatchCityLoc

WACV 2027

MatchCityLocAerial Localization in LoD City Models
via Surface Normal Matching

Oussema Dhaouadi1,2,3,4 Zuria Bauer3 Johannes Meier2,4 Marc Pollefeys3,5 Daniel Cremers2,4 Olaf Wysocki1
1 University of Cambridge 2 TUM 3 ETH Zürich 4 MCML 5 Microsoft
Paper Code Video Poster Datasets
Paper teaser: prior methods score predicted and rendered building silhouettes; MatchCityLoc matches predicted and rendered surface normals, then refines the pose with PnP.
From silhouette scoring to explicit 3D pose solving. Prior approaches score many rendered masks. MatchCityLoc matches surface normals and refines the pose with PnP. Paper figure ↗

Abstract

Autonomous Unmanned Aerial Vehicle (UAV) navigation in urban environments requires precise pose estimation. Because Global Navigation Satellite System (GNSS) signals are often unreliable in cities, visual localization serves as a crucial complement. Within this domain, textureless Level-of-Detail (LoD) city models provide a lightweight and privacy-preserving substitute for heavy 3D textured meshes. However, localizing against abstract geometry is challenging. Existing pipelines typically score 2D silhouette overlaps, which can suffer from severe geometric ambiguities.

We present MatchCityLoc, a framework that shifts this paradigm to direct feature matching from a noisy 6-DoF pose prior. By extracting surface normals from the query image using a zero-shot foundation model and matching them against rendered LoD normals, we establish explicit 3D-to-2D correspondences. This enables a Perspective-n-Point (PnP) pose solve, which we iteratively refine through a render-and-match loop. Experiments demonstrate strong pose recall and accurate translation and rotation among the evaluated off-the-shelf LoD-based methods with very few renderings. By bridging the modality gap between aerial imagery and texture-free geometry, MatchCityLoc enables accurate localization using lightweight, privacy-preserving city models.

Method

Surface normals provide a common representation for the camera image and the city model. Refinement starts from a noisy position-and-orientation prior with known camera intrinsics.

Click a block to explore the pipeline
Fixed query normals → matching
Query normals stay fixed↶ Re-render at the updated pose · repeat steps 2–4
01 / 04

A shared representation

MoGe-2 predicts surface normals from the UAV image. This prediction is computed once and reused throughout refinement: matching uses surface orientation rather than RGB texture.

Input
UAV image
Output
Fixed query normal map

Iterative pose refinement

Step through the recorded experiment used in the paper’s refinement figure.

1.8 s per step · repeats
Fixed query surface-normal prediction

Predicted normals Fixed query

LoD normals rendered at the initial pose

Rendered normals Initial pose

LoD edges overlaid on the UAV photograph at the initial pose

Reprojection onto the image Red: projected LoD edges

Enable JavaScript to view the recorded matches.

Initial poseTranslation error 19.8 mRotation error 1.9°

Qualitative results

Pose estimates reprojected into real UAV images across all three benchmarks.

Original paper panels, including measured errors. LoD-Loc v2 and SemCityLoc use off-the-shelf masks; LoD-Loc v3 uses oracle masks. Red lines are projected LoD edges, not matches.

Quantitative results

Tables follow the paper's order and comparison groups. Click a heading to sort within each group, reverse, then restore paper order.

Median pose error and recall from Tables 1–2.
DatasetSplit / modelTranslation ↓Rotation ↓R @ 2 m / 2° ↑R @ 5 m / 5° ↑
UAVD4L-LoDin-Traj · LoD30.62 m0.34°92.64%98.44%
UAVD4L-LoDout-of-Traj · LoD30.66 m0.28°94.66%99.22%
Swiss-EPFLin-Place · LoD22.02 m1.64°45.51%74.16%
Swiss-EPFLout-of-Place · LoD20.99 m1.16°72.82%89.71%
SemCityLockeDLoD21.45 m1.14°59.12%77.16%
SemCityLockeDLoD30.87 m0.81°70.33%86.34%

Joint recall requires both translation and rotation errors to fall within the stated thresholds. Our SemCityLockeD evaluation uses 937 rectified queries. Published results marked ‡ use their original protocol with mean translation and yaw errors, unlike our median translation and full-rotation errors, and are excluded from the best/second-best highlighting.

Citation

@inproceedings{dhaouadi2027matchcityloc,
  title     = {{MatchCityLoc}: Aerial Localization in {LoD} City Models via Surface Normal Matching},
  author    = {Dhaouadi, Oussema and Bauer, Zuria and Meier, Johannes and Pollefeys, Marc and Cremers, Daniel and Wysocki, Olaf},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027},
  url       = {https://arxiv.org/abs/2610.27110}
}