We present LoD-Loc v3, a novel method for generalized aerial visual localization in dense urban environments. While prior work LoD-Loc v2 achieves localization through semantic building silhouette alignment with low-detail city models, it suffers from two key limitations: poor cross-scene generalization and frequent failure in dense building scenes. Our method addresses these challenges through two key innovations. First, we develop a new synthetic data generation pipeline that produces InsLoD-Loc - the largest instance segmentation dataset for aerial imagery to date, comprising 100k images with precise instance building annotations. This enables trained models to exhibit remarkable zero-shot generalization capability. Second, we reformulate the localization paradigm by shifting from semantic to instance silhouette alignment, which significantly reduces pose estimation ambiguity in dense scenes. Extensive experiments demonstrate that LoD-Loc v3 outperforms existing state-of-the-art (SOTA) baselines, achieving superior performance in both cross-scene and dense urban scenarios with a large margin. The project is available at https://nudt-sawlab.github.io/LoD-Locv3/.
Figures & tables
Figure 2: Visualization of instanced LoD models and labels. The top row displays the instanced LoD models. The bottom row illustrates the rendered instance labels rendered from the models, alongside their corresponding semantic labels.
Figure 3: Overview of the InsLoD-Loc dataset. The left panel illustrates the geographic distribution of the 40 flight areas across Europe and Asia. The right panel showcases representative samples from the dataset, each displaying (from left to right): a photorealistic RGB query image and its corresponding pixel-accurate instance label, where each color represents a unique building.
Camera
Camera Configuration
Sampling Methodology
Resolution[px]
Sensor Size
FOV
Strategy
Path
View Angle
Height[m]
Camera 1
1600×1200
16mm×12mm
45°
Sequence
Irregular
[0°, 70°]
[200, 500]
Camera 2
1920×1080
23.76mm×13.365mm
25°
Grid
Uniform
0°, 45°
[200, 500]
Camera 3
1920×1080
23.76mm×13.365mm
25°
Sequence
Regular
0°, 45°
[400, 500]
Table 1: Camera configurations and sampling methodologies used for data acquisition.
Method
in-Traj.
out-of-Traj.
2m-2°
3m-3°
5m-5°
T.e./R.e.
2m-2°
3m-3°
5m-5°
T.e./R.e.
Prior
0
0
4.30
6.48/1.63
0
0
0.36
11.10/0.92
CAD-Loc [ 65 ]
SIFT+NN
0
0
0
-
0
0
0
-
SPP+SPG
0
0
0
-
0
0
0
-
LoFTR
0
0
0
-
0
0
0
-
e-LoFTR
0
0
0
-
0
0
0
-
Table 2: Quantitative comparison results of different methods over UAVD4L-LoDv2 dataset. T.e. and R.e. denote median translation error (m) and median rotation error (°). † indicates models trained in-distribution on this dataset. Our method utilizes area-based weighting.
Method
in-Place.
out-of-Place.
2m-2°
3m-3°
5m-5°
T.e./R.e.
2m-2°
3m-3°
5m-5°
T.e./R.e.
Prior
0
0
0.56
17.6/3.87
0
0
1.06
17.9/3.94
CAD-Loc
same*
0
0
0
-
0
0
0
-
MC-Loc
DINOv2
0.90
4.40
17.50
8.18/2.54
2.90
9.00
30.20
6.23/1.97
RoMa
0.20
1.20
4.80
9.80/2.65
0.70
2.10
11.5
10.3/3.97
LoD-Loc†
-
36.79
50.56
69.77
2.87/1.78
14.24
31.39
59.89
8.73/2.78
Table 3: Quantitative comparison results of different methods over Swiss-EPFLv2 dataset. The same* indicates that the statistics are identical to those in Tab. 2 . †, T.e., and R.e. have the same meanings as those in Tab. 2 . Our method utilizes area-based weighting.
Method
Grid-Traj.
Sequence-Traj.
2m-2°
3m-3°
5m-5°
T.e./R.e.
2m-2°
3m-3°
5m-5°
T.e./R.e.
Prior
0.10
1.70
8.90
8.03/1.78
0.30
0.80
8.80
8.19//1.76
CAD-Loc
same*
0
0
0
-
0
0
0
-
MC-Loc
DINOv2
0
0.26
1.61
31.95/1.88
3.42
10.50
29.81
5.16/0.49
RoMa
0
0.14
0.39
32.36/1.89
3.58
10.72
31.12
5.16/0.48
LoD-Loc
-
10.52
20.95
33.80
8.03/1.78
0
0
5.06
8.19/1.76
Table 4: Quantitative comparison results of different methods over Tokyo-LoDv3 dataset. The same* indicates that the statistics are identical to those in Tab. 2 . T.e. and R.e. have the same meanings as in Tab. 2 . LoD-Loc v3 c method utilizes confidence-based weighting, and LoD-Loc v3 a method utilizes area-based weighting.
Figure 4: Visualization of localization results on the Tokyo-LoDv3 dataset. Superimposed instance masks rendered at estimated poses demonstrate that our method effectively resolves ambiguity in dense urban scenes. The columns from left to right show: query image, prior pose, LoD-Loc v2, LoD-Loc v3 (Ours), and ground truth.
Dataset
Method
in (Grid)-Traj. (Place.)
out-of (Sequence)-Traj. (Place.)
2m-2 ∘
3m-3°
5m-5°
T.e./R.e.
2m-2°
3m-3°
5m-5°
T.e./R.e.
UAVD4L-LoDv2
LoD-Loc v2
71.40
91.20
96.10
1.66/0.34
70.00
89.80
96.10
1.62/0.39
LoD-Loc v3
97.60
98.90
99.70
0.49/0.13
97.40
99.00
99.40
0.60/0.12
Swiss-EPFLv2
LoD-Loc v2
11.90
28.10
52.00
4.79/1.42
24.00
49.30
77.80
3.02/1.12
LoD-Loc v3
58.60
79.90
95.40
1.61/0.77
36.90
64.10
88.90
2.43/0.83
Tokyo-LoDv3
LoD-Loc v2
22.70
43.80
74.70
2.91/0.32
35.60
65.70
92.00
2.41/0.27
Table 5: Ablation on Silhouette Representation. Quantitative comparison of LoD-Loc v2 (semantic) and LoD-Loc v3 (instance) trained on the same InsLoD-Loc dataset.
Alignment
Grid - Traj.
Sequence - Traj
2m-2°/3m-3°/3m-3°
2m-2°/3m-3°/3m-3°
LoD-Loc v2
19.60/39.40/72.10
21.50/47.80/89.00
LoD-Loc v3
38.10/65.40/86.40
49.80/79.90/95.80
Table 6: Ablation on Instance Alignment. Quantitative comparison of localization performance using separated instance masks versus artificially merged semantic masks.
Technical University of Munich, Munich, Germany · CV4DT, University of Cambridge, Cambridge, United Kingdom · Munich Center for Machine Learning, Munich, Germany +3