GOOSELoc: A Multimodal Benchmark for Ground-to-Aerial Localization in Unstructured Off-Road Terrain
Authors: Fabian Deuser, Peter Mortimer, Robin-Nico Kampa, Anna Bößendörfer, Mirko Mählisch, Martin Werner, Norbert Oswald
Organizations: Institute of Distributed Intelligent Systems (ETTI 2), University of the Bundeswehr Munich, Neubiberg, Germany · Technical University of Munich, Germany; TUM School of Engineering and Design, Department of Aerospace and Geodesy · Institute for Autonomous Driving (LRT 8.1), University of the Bundeswehr Munich, Neubiberg, Germany · TUM School of Computation, Information and Technology; Munich Data Science Institute (MDSI)
Cross-view pose estimation estimates a ground camera's geographic pose by matching onboard sensor data with geo-referenced overhead imagery. This is particularly relevant when GNSS is unreliable, for example under forest canopy or in urban canyons. Existing metric cross-view benchmarks, such as VIGOR and cross-view KITTI, are dominated by urban and road scenes with strong human-made structure without off-road terrain. We introduce GOOSELoc, a multimodal benchmark for ground-to-aerial localization in unstructured environments, including forests, fields, and rural tracks. GOOSELoc comprises 9,356 scenes recorded across all four seasons around Munich, Germany, pairing geo-referenced aerial orthophotos with surround-view RGB imagery and 3D LiDAR point clouds. Paved/unpaved labels enable analyzing the impact of terrain structure, while the seasonal recordings allow studying the appearance gap to mostly summer-captured orthophotos. To assess the transferability of existing methods, we adapt and fine-tune four state-of-the-art cross-view localization baselines on GOOSELoc. Despite task-specific fine-tuning, the performance degrades substantially in off-road environments. We therefore propose a multimodal baseline combining surround-view imagery and LiDAR geometry, improving localization performance and providing a stronger reference for future work. Data and baselines are released at https://gooseloc.github.io.
Figures & tables
Figure 1: GOOSELoc Dataset Sample. The 4 roof-mounted cameras ■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■ each have a 2448×1000 px resolution and a 94.71\SIUnitSymbolDegree horizontal field of view after rectification, giving us a 360\SIUnitSymbolDegree coverage of the surroundings. This can be fused with a sweep of the roof-mounted 128-beam rotating LiDAR scanner ■■■■■■■■■■■■■■■■■ . The Bavarian orthophotos ■■■■■■■■■■■■■■■■■ are taken on a clear day in the summer and have a temporal mismatch to the ground-view images recorded during winter. The ground truth global location and heading ★★★★★★★★★★★★★★★★★ are obtained by an Oxford RT3000v3 INS system with differential RTK-GNSS corrections using the NTRIP protocol to receive accurate readings in forested areas.
Dataset
Size
Ground view
Modalities
Environment
Seasonal gap
Localization
CVUSA [ 36 , 41 ]
44.4k
Panorama
RGB
Urban / rural
No
Retrieval
CVACT [ 16 ]
137.2k
Panorama
RGB
Urban
No
Retrieval
VIGOR [ 46 ]
105.2k
Panorama
RGB
Urban
No
Retrieval + Metric
Cross-View KITTI [ 26 ]
30.2k
Forward
RGB
Urban / rural
No
Metric
University-1652 [ 43 ]
146.0k
Social Media
RGB
Urban (buildings-only)
No
Retrieval
MoViX / TartanDrive 2.0 [ 3 , 30 ]
4.6k
Forward (video)
RGB
Off-road
Yes
Metric
Table 1: GOOSELoc in the context of existing cross-view benchmarks. A GOOSELoc scene comprises four surround images, a LiDAR sweep, and a geo-referenced orthophoto crop. GOOSELoc is the only benchmark pairing surround-view imagery and LiDAR against official orthophotos, and the only one combining paved-to-off-road coverage with a controlled seasonal ground-aerial gap. McPed [ 40 ] is the closest multimodal relative: RGB + LiDAR from a single forward-facing 79∘ camera on a tracked robot, recorded in clear weather without a seasonal protocol, with SLAM-derived ground truth and per-sequence rather than geographic splits. MoViX shares the off-road setting and a ground-aerial temporal gap but relies on a single forward-facing RGB camera.
Table 2: Composition of GOOSELoc. Scene counts per split by season and terrain. Scenes that are not paved are unpaved ( 45.8% paved overall).
Figure 3: The two heading settings on one scene ( ⋆ = GT position, red = true heading, white dashed = reference “up”, cyan = admissible range of Δθ ). Left – north-up: no prior, recover the full 360∘ . Right – IMU prior: ±10∘ coarse prior, recover only the residual ∣Δθ∣≤10∘ (offset sampled uniformly in [−10∘,+10∘] at train and test).
Figure 4: Our architecture. Frozen DINOv3 and Utonia yield a shared token set. In the position branch each modality correlates its own query with the DOP20 patch features, and an evidence gate weights the two resulting score maps before they are summed. The orientation branch correlates the camera and LiDAR-wedge tokens against the tile and feeds the match maps, with the camera [CLS] tokens, to a 72-bin heading classifier.
dev
test
↓ Localization (m)
↓ Orientation ( ∘ )
↓ Localization (m)
↓ Orientation ( ∘ )
Setting
Methods
Mean
Median
Mean
Median
Mean
Median
Mean
Median
North-up
Predict center
38.9
40.9
90.0
90.0
39.7
41.5
90.0
90.0
CCVPE (surround)
33.6 ± 0.4
25.7 ± 1.3
37.2 ± 8.5
16.7 ± 12.2
45.4 ± 2.9
37.0 ± 2.9
52.8 ± 5.0
35.6 ± 17.0
HC-Net (surround)
31.3 ± 0.2
27.3 ± 0.6
85.7 ± 1.5
84.2 ± 4.8
31.8 ± 0.4
28.6 ± 0.2
87.0 ± 0.5
84.4 ± 1.2
FG 2 (front)
32.8 ± 0.9
29.3 ± 0.2
89.9 ± 0.4
91.3 ± 0.4
37.7 ± 1.2
35.1 ± 0.8
92.0 ± 0.2
93.4 ± 0.1
Table 3: Baseline localization on GOOSELoc (dev and test). Mean and median error for position (m) and orientation ( ∘ ), averaged over three seeds. Predict center is a non-learned trivial baseline. We evaluate FG 2 on the native front camera and on our surround-view adaptation, CCVPE and HC-Net on our surround pseudo-panorama, and Shi et al. on the front camera (its native configuration). CCVPE and HC-Net appear only in the north-aligned setting, Shi et al. only under the IMU prior (bounded-prior method). Best per column within each setting in bold . The IMU-prior setting predicts a bounded residual ∣Δθ∣≤10∘ .
Paved
Unpaved
Setting
Method
Loc (m)
Ori ( ∘ )
Loc (m)
Ori ( ∘ )
North-up
FG 2 (surround)
27.6 ± 0.4
89.9 ± 0.2
36.1 ± 0.4
98.0 ± 0.9
Ours (RGB)
25.5 ± 0.7
28.6 ± 3.3
25.5 ± 0.3
27.6 ± 4.2
Ours (LiDAR-only)
23.2 ± 0.6
35.6 ± 7.3
26.5 ± 0.9
40.8 ± 3.9
Ours (RGB+L)
21.2 ± 0.3
24.8 ± 1.6
22.9 ± 0.7
25.3 ± 1.1
IMU ±10∘
FG 2 (surround)
13.8 ± 0.1
4.0 ± 0.2
27.0 ± 0.7
4.7 ± 0.2
Table 4: Performance by terrain (test, median ± std over 3 seeds).
Setting
Winter
Spring
Summer
Autumn
/ Method
Loc(m)
Ori( ∘ )
Loc(m)
Ori( ∘ )
Loc(m)
Ori( ∘ )
Loc(m)
Ori( ∘ )
North-up
FG 2
28.4 ± 1.2
80.3 ± 0.8
28.5 ± 0.7
90.2 ± 0.9
36.5 ± 0.6
100.1 ± 1.5
34.6 ± 0.4
108.5 ± 0.6
RGB
25.5 ± 0.7
26.9 ± 2.9
24.3 ± 0.1
36.5 ± 10.2
26.3 ± 0.6
28.3 ± 1.2
25.9 ± 0.2
19.5 ± 2.7
L-only
23.9 ± 1.1
44.4 ± 4.4
23.5 ± 1.2
25.7 ± 4.8
26.1 ± 0.7
44.8 ± 1.4
26.7 ± 1.5
45.3 ± 15.1
RGB+L
22.0 ± 1.0
25.3 ± 0.7
20.3 ± 0.5
30.8 ± 0.1
24.1 ± 0.5
28.8 ± 2.3
22.7 ± 0.8
16.3 ± 1.4
IMU
FG 2
16.9 ± 0.6
4.6 ± 0.1
14.2 ± 0.1
4.0 ± 0.2
26.6 ± 1.1
4.8 ± 0.2
25.1 ± 1.0
4.5 ± 0.3
Table 5: Performance by season (test, median ± std over 3 seeds).
Figure 5: Qualitative predictions on GOOSELoc (test, north-up, no prior). Gold ⋆ / arrow: ground-truth position / heading, red ∘ /arrow: prediction; the viridis overlay is the predicted position posterior.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Split
Scenes
GNSS position mode (%)
RTK Integer (cm)
RTK Float (dm)
Other
train
6430
74.4
10.4
15.2
dev
1064
89.6
5.9
4.5
test
1862
70.4
9.0
20.6
Appendix
Table A1: Ground truth pose quality per split, decoded from the gps_pos_mode byte of the NCOM packets; rows sum to 100% . “Other” groups the status codes Differential , SPS and packets with no position solution at all. Pose quality was not a split constraint, but care was taken to maintain a good balance for the train and test split.
Table A2 : Robustness to heading-prior error (test, median; Loc in m, Ori in ∘ ; seed-0 checkpoints). As the injected prior error ∣Δθ∣ grows, localization stays nearly flat up to ±20∘ and loses only ∼3 m at ±45∘ ; orientation stays at a few degrees up to ±10∘ and degrades gracefully beyond ( 0∘ = perfect prior). A coarse prior suffices, and the fusion advantage persists at every prior quality.
Paved
Unpaved
Look-alike positions ( τ=0.8 )
79 ( 1.9% )
101 ( 2.5% )
Look-alike fraction ( τ=0.7 )
5.6%
7.8%
Appendix
Table A3 : Perceptual aliasing by terrain (test, per-scene medians, unpaved > paved). Unpaved scenes contain 1.28× more look-alike positions at τ=0.8 .
Winter
Spring
Summer
Autumn
Look-alike positions ( τ=0.8 )
63
79
117
102
Appendix
Table A4 : Perceptual aliasing by season (median per-scene look-alike count). Aliasing peaks in the leaf-on seasons, matching where the appearance-based baselines degrade most.
Method
Setting
ρ (all)
ρ (paved)
ρ (unpaved)
FG 2 (surround)
IMU
+0.23
+0.18
+0.20
Shi et al. (front)
IMU
+0.21
+0.19
+0.20
Ours (RGB+L)
IMU
+0.18
+0.26
+0.05
CCVPE (surround)
NA
+0.16
+0.17
+0.10
Ours (LiDAR-only)
IMU
+0.15
+0.23
+0.01
FG 2 (surround)
NA
+0.13
+0.08
+0.10
Appendix
Table A5 : Rank correlation between per-scene aliasing and localization error (Spearman ρ , look-alike count at τ=0.8 , averaged over training seeds). Every method correlates positively. The matching-based baselines correlate most and remain positive within both terrain groups. Our fusion is least affected in the prior-free setting.
Capture date
train
dev
test
2024-06-18
550
72
3
2024-07-31
1590
177
326
2024-08-24
3920
749
1517
2025-08-19
27
3
16
2025-08-26
310
37
0
Appendix
Table A6 : DOP20 orthophoto capture dates per split. All survey flights fall in the leaf-on season. The two most recent dates stem from the 2025 flight campaign of the same areas.
Full test ( n=1862 )
Buffered ( n=840 )
Setting
Method
Loc (m)
Ori ( ∘ )
Loc (m)
Ori ( ∘ )
North-up
CCVPE (surround)
45.4 ± 2.9
35.5 ± 17.0
46.3 ± 2.8
39.9 ± 13.0
HC-Net (surround)
31.8 ± 0.4
84.4 ± 1.2
31.8 ± 0.2
83.6 ± 2.1
FG 2 (surround)
34.4 ± 0.4
93.1 ± 0.3
34.2 ± 0.2
92.8 ± 0.2
Ours (RGB)
29.8 ± 0.1
27.9 ± 3.3
28.8 ± 0.1
35.1 ± 10.6
Ours (LiDAR-only)
28.7 ± 0.3
38.5 ± 5.4
29.2 ± 0.4
49.3 ± 6.4
Appendix
Table A7 : Full vs. buffered test (Loc = mean position error, Ori = median orientation error, ± std over 3 seeds). No method gains from the removed overlap: every ranking persists on the leakage-free subset, the paired fusion advantage grows (cf. the main text), and orientation degrades most for the camera-free variant.
Pair
Min
5th pct.
Median
Tile overlap
test → train
1.5 m
30 m
252 m
40%
dev → train
0.4 m
20 m
179 m
58%
test → dev
0.1 m
10 m
141 m
66%
Appendix
Table A8 : Inter-split nearest-neighbor distances between scene positions, and the share of scenes whose 204.8 m aerial crop overlaps a crop of the other split.
Cross-view geolocalization aims to determine the geographic location of a ground-level image by retrieving its corresponding satellite image, offering a promising alternative for localization in Global Navigation Satellite System (GNSS)-denied or degraded environments such as urban canyons, dense forests, and disaster scenarios. Despite recent progress, existing approaches remain limited by the lack of globally diverse training datasets and benchmarks that are often geographically restricted. Furthermore, most benchmarks rely on panoramic ground views, which do not reflect the limited field-of-view imagery captured by real world cameras. To address these challenges, we introduce Cross-View World 500k (CVW500k), a large-scale globally distributed dataset containing over 500,000 ground-satellite image pairs collected across diverse geographic regions and environments. Unlike previous datasets, CVW500k uses limited field-of-view ground images, providing a more realistic and challenging setting for cross-view geolocalization. In addition, we propose GeoQueryNet, a query-based cross-view fusion transformer framework designed to bridge the viewpoint gap between ground and satellite images. The model leverages a CLIP-based visual encoder with parameter-efficient LoRA adaptation and introduces a Cross-View Alignment Module that uses learnable query tokens to extract geographically consistent features across viewpoints. Our code can be found here: https://github.com/Fahim17/Cross_View_World_Dataset.git .
Fahimul Aleem, Shruti Vyas
University of Central Florida, Orlando, FL, United States
We present M3GA-Wild, the first benchmark for multi-modal, multi-session ground-to-aerial place recognition in forests. M3GA-Wild unifies and extends existing forest localisation datasets, providing a holistic benchmark with synchronised RGB imagery and LiDAR from ground traversals spanning 36 km, aligned high-resolution aerial imagery and multi-altitude LiDAR covering 370 hectares, and accurate geo-referenced 6-DoF poses for precise evaluation. M3GA-Wild captures diverse forest scenes with varying viewpoints, occlusion, and environmental conditions, enabling systematic evaluation of visual, LiDAR, cross-modal, and multi-modal methods. Baseline experiments show that LiDAR-based approaches significantly outperform vision-only methods under severe viewpoint differences, while current multi-modal fusion strategies yield limited gains due to poor cross-modal alignment. By pairing aerial RGB imagery with geo-referenced aerial LiDAR, M3GA-Wild also enables evaluation of foundation models for monocular depth estimation as a cheap source of 3D geometry from forest imagery, with initial experiments revealing shortfalls of current methods. These results highlight key challenges in cross-platform localisation, including modality misalignment and severe domain gaps. M3GA-Wild establishes a new benchmark to support research in robust multi-modal localisation and long-term autonomy in unstructured natural environments. The dataset and code will be available upon acceptance.
Ethan Griffiths, Maryam Haghighat, Simon Denman +2
Queensland University of Technology (QUT) · CSIRO Robotics
Consistent cross-view understanding under extreme viewpoint changes is essential for spatial intelligence, as it enables models to recognize the same scene across extreme viewpoint gaps. Cross-view localization naturally provides a promising pathway toward this ability, as it requires a model to align ground-view imagery with geo-referenced satellite-view imagery despite drastic appearance changes to estimate camera poses. Recent visual foundation models have made this long-standing localization problem increasingly feasible by providing rich 2D representations for cross-view matching. However, we argue that cross-view localization should not be viewed merely as 2D matching or pose estimation. In this work, we revisit cross-view localization as more than pose estimation and investigate how it can help the model develop consistent cross-view understanding under extreme viewpoint changes, including stable semantics, reliable structure, and transferable geometry. We identify three key limitations of existing methods that prevent them from achieving this. They usually lack explicit 3D grounding, rely on strict point-wise matching that can weaken semantic consistency, and learn from an absolute objective that provides limited guidance for geometric reasoning. To address these limitations, we propose CROSS, a unified cross-view localization framework built upon 3D-grounded alignment, structure-aware matching, and hypothesis ranking. This formulation makes structure learning an intrinsic requirement, encourages semantic representations to remain stable, and enables the model to acquire transferable geometry. Extensive experiments on the KITTI and VIGOR datasets show that CROSS achieves state-of-the-art performance in cross-view localization. More importantly, CROSS effectively learns stable semantics, reliable structure, and transferable geometry across extremely different viewpoints.
Mao Chen, Xiangkai Zhang, Zhiyong Liu +2
State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, and also with the School of Artificial Intelligence, University of Chinese Academy of Sciences · Beijing Aerospace Control Center