GOOSELoc: A Multimodal Benchmark for Ground-to-Aerial Localization in Unstructured Off-Road Terrain
Paper: https://huggingface.co/papers/2611.16162 | Project: https://gooseloc.github.io
WACV 2027
Fabian Deuser, Peter Mortimer, Robin-Nico Kampa, Anna Bößendörfer, Mirko Mählisch, Martin Werner, Norbert Oswald
University of the Bundeswehr Munich and Technical University of Munich
This is the data release (v1.0) of GOOSELoc: 9,356 scenes recorded around Munich across all four seasons, each with four surround-view RGB images, one 128-beam LiDAR sweep and a geo-referenced 20 cm DOP20 orthophoto crop. It also holds the checkpoints of all methods of the paper.
The data loader, the official evaluation and the code of all baselines are at https://github.com/GOOSELoc/GOOSELoc.
A winter test scene under forest canopy. Gold: ground-truth position and heading; blue: the LiDAR sweep at the ground-truth pose. Aerial imagery: Datenquelle: Bayerische Vermessungsverwaltung – www.geodaten.bayern.de (CC BY 4.0).
Abstract
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.
Download
Clone the code repository and download the data into data/GOOSELoc inside
it, where the code finds it by default:
git clone https://github.com/GOOSELoc/GOOSELoc.git && cd GOOSELoc
pip install -e .
pip install -U huggingface_hub
hf download Skyy93/GOOSELoc --repo-type dataset --local-dir data/GOOSELoc
gooseloc-check --checksums
To download only some parts, add --include:
| part | size | --include |
|---|---|---|
| metadata, checksums, README, license | < 10 MB | "metadata/*" "CHECKSUMS.sha256" "README.md" "LICENSE.md" |
| aerial crops | 4.5 GB | "aerial/*" |
| LiDAR sweeps | 26 GB | "lidar/*" |
| cameras (all four) | 40 GB | "cameras/*" (or "cameras/front/*") |
| checkpoints | 4.1 GB | "checkpoints/*" (or "checkpoints/<method>/*") |
Layout
GOOSELoc/
├── README.md, LICENSE.md, CHECKSUMS.sha256
├── metadata/
│ ├── scenes.csv one row per scene: split, ground truth, protocol
│ ├── splits/{train,dev,test,test_buffered}.txt
│ └── calibration.yaml camera intrinsics, sensor frames (preliminary)
├── aerial/<scene_id>.jpg 1024 x 1024 RGB, 0.20 m/px, north-up
├── cameras/{front,right,back,left}/<scene_id>.jpg 2448 x 1000 RGB
├── lidar/<scene_id>.bin float32 (N, 4): x, y, z, intensity
└── checkpoints/{gooselocnet,fg2,ccvpe,hcnet,shi}/ checkpoints of the paper
Images are JPEG (quality 95, no chroma subsampling). The scene id is the
capture timestamp in nanoseconds (UTC). CHECKSUMS.sha256 covers the
sensor files and metadata/; each checkpoint folder has its own
MANIFEST.json with SHA-256 sums and training details.
| split | scenes | paved | unpaved | winter | spring | summer | autumn |
|---|---|---|---|---|---|---|---|
| train | 6,430 | 2,988 | 3,442 | 1,387 | 1,976 | 1,341 | 1,726 |
| dev | 1,064 | 400 | 664 | 154 | 304 | 103 | 503 |
| test | 1,862 | 900 | 962 | 332 | 608 | 532 | 390 |
| test_buffered (subset of test) | 840 |
The splits assign whole 1 km x 1 km cells to train, dev or test. The test split
keeps only scenes whose GNSS fix and INS pose agree within 5 m.
test_buffered additionally drops every test scene whose aerial crop could
share a pixel with a training crop.
Essentials
- Task: given the ground observation (cameras and/or LiDAR) and the aerial crop, predict the vehicle position in the crop and its heading.
- Ground truth:
gt_px_x,gt_px_yinscenes.csvare the vehicle position in pixels of the north-up crop (origin top-left, x right/east, y down/south).heading_degis the compass heading (clockwise from north). The vehicle is placed uniformly at random in the central 512 x 512 px. Poses come from an Oxford RT3000v3 INS, RTK-corrected where available and otherwise from its raw GNSS fix (see the appendix of the paper). - Settings: in north_up the crop is used as is. In imu_prior the crop is
rotated counter-clockwise by
heading_deg + imu_prior_offset_deg, where the offset is the fixed per-scene prior error (uniform in ±10°) used for dev and test. - Georeference:
tile_utm_e0,tile_utm_n0are the UTM 32N (EPSG:25832) coordinates of the crop's top-left corner;utm = (e0 + 0.2 x, n0 - 0.2 y).gnss_lat,gnss_lonhold the raw GNSS fix as a realistic position prior. - LiDAR: vehicle frame, x forward, y left, z up, in metres. 128-beam Velodyne, one sweep per scene.
- Cameras: four roof-mounted Basler cameras at 0°, 90°, 180° and 270° clockwise from the vehicle's forward axis, cropped to 2448 x 1000 px.
- Aerial: crops of the Bavarian DOP20 orthophotos (20 cm/px), flown June–August 2024 and August 2025, so every scene has a temporal gap to its aerial image.
The full column reference and all conventions are in docs/DATASET.md of the code repository.
License
The dataset combines two sources under different licenses. Each part keeps its own license; there is no single license for the release as a whole. See LICENSE.md for the full terms.
| part | license | attribution |
|---|---|---|
ground data (cameras/, lidar/) and metadata/ |
CC BY-SA 4.0, like the GOOSE dataset they are derived from | cite GOOSELoc and GOOSE |
aerial crops (aerial/) |
CC BY 4.0, like the DOP20 source data | Datenquelle: Bayerische Vermessungsverwaltung – www.geodaten.bayern.de |
checkpoints (checkpoints/) |
the license of the method's code: GOOSELocNet and Shi et al. MIT, FG² and CCVPE GPL-3.0, HC-Net see below | cite the method |
- Modified versions of the ground data have to be shared under CC BY-SA 4.0.
- Any use of the aerial crops, including figures, has to carry the attribution above.
- The aerial crops were cut from the DOP20 source tiles, composited where a crop straddles tile borders and JPEG-encoded; they are otherwise unchanged.
- The HC-Net upstream code has no license file. Its checkpoints are provided for reproducing the paper's results.
- GOOSELocNet uses frozen DINOv3 and Utonia encoders. Their weights are not part of this release; they are downloaded separately under their own licenses.
Citation
@inproceedings{deuser2027gooseloc,
title = {{GOOSELoc}: A Multimodal Benchmark for Ground-to-Aerial Localization in Unstructured Off-Road Terrain},
author = {Deuser, Fabian and Mortimer, Peter and Kampa, Robin-Nico and B{\"o}{\ss}end{\"o}rfer, Anna and M{\"a}hlisch, Mirko and Werner, Martin and Oswald, Norbert},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
year = {2027}
}
@inproceedings{mortimer2024goose,
title = {The {GOOSE} Dataset for Perception in Unstructured Environments},
author = {Mortimer, Peter and Hagmanns, Raphael and Granero, Miguel and Luettel, Thorsten and Petereit, Janko and Wuensche, Hans-Joachim},
booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
year = {2024}
}
GOOSELoc builds on the GOOSE dataset recorded with the MuCAR-3 vehicle of the Institute for Autonomous Driving (UniBw Munich).
- Downloads last month
- 133
