PlanarCityLoc
WACV 2027

PlanarCityLocBreaking Silhouette Ambiguity in LoD Aerial Localization via 3D Planes

We combine silhouette-based coarse initialization with explicit 3D–2D plane correspondences to refine camera pose in lightweight, textureless LoD city models.

Paper Code Poster Video Datasets

3D planar grounding

Rotate the city model, select a plane, and move the camera.

3D city model

Drag to rotate · select a surface

Camera projection

Same selected plane, projected into the image
Solid: current poseDashed: reference pose

Loading the LoD model…

Similar 2D silhouettes can occur at different camera poses. Move the camera to inspect the ambiguity.

−12 mReference pose+12 m

The slider translates the camera along its horizontal axis. It does not rotate the model.

Conceptual score landscape

LoD-Loc v2LoD-Loc v3Ours
0.2.4.6.8−40−2002040 Pose error (m / °) Score Ambiguous overlap Zero gradient
Interactive illustration: select a 3D plane to see its image projection, then vary the camera pose. Camera motion and score curves are illustrative. Model data · Data sources and terms.
01 / Method

The PlanarCityLoc pipeline

Follow the query and the city model through geometry extraction, coarse initialization, and iterative 3D planar grounding.

Pipeline overviewa / e
0:00 / 0:25
(a) Query geometry

Geometry and semantics extraction

The camera-ready configuration uses MoGe-2 ViT-S for depth and normals; SAM 3 supplies a building mask. Sobel filtering of the normals provides structural edges.

Input
RGB query image Iq
Output
Depth Dq, normals Nq, mask Mq, edges Eq

Animated algorithm overview. Query geometry stays fixed while plane correspondences refine the camera pose.

02 / Qualitative evaluation

Visual comparisons

Compare 10 selected examples across two datasets. The gallery shows cases where PlanarCityLoc has lower translation and rotation errors than the methods available for that example; these per-image examples are not aggregate results.

Same query and image coordinates on both sides. Red lines are LoD edges rendered at the selected pose. Use the arrow keys to move the divider; Home/End reveal one method completely.

Errors: translation (m) / rotation (°). Examples compare LoD-Loc v2 + SAM 3, the particle-filter stage, and PlanarCityLoc. Saved-run settings can differ from the aggregate benchmark. Additional prediction details · Swiss model: ©swisstopo.

03 / Quantitative evaluation

Quantitative results

Paper order · Click a heading to sort within groups.

UAVD4L-LoDv2 · in-Traj · LoD3

Joint-threshold recall (%) and median translation/rotation errors. Best and second-best distinct values are bold and underlined within each setup group. † Adapted LoD-Loc evaluations with nvdiffrast rendering and either ground-truth-rendered masks or off-the-shelf SAM 3 predictions. * Uses the InsLoD-Loc training set.

Citation

@inproceedings{dhaouadi2027planarcityloc,
  title     = {{PlanarCityLoc}: Breaking Silhouette Ambiguity in {LoD} Aerial Localization via 3{D} Planes},
  author    = {Dhaouadi, Oussema and Hu Chen, Shao Jie and
               Gollreiter, Jakob and Jusufi, Samir and Rieger, Nikolas and
               Wendlandt, Caroline and Wender, Christian and Mao, Jingfeng and
               Wang, Guangming and Bauer, Zuria and Pollefeys, Marc and
               Cremers, Daniel and Wysocki, Olaf},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}