You need to agree to share your contact information to access this dataset
Dataset packaging, coordinate alignment, and licensing terms are currently undergoing active verification. Access will be reviewed and granted after verification is complete for authorized academic researchers.
The dataset, image permissions, and licensing terms are currently undergoing active verification. Please specify your affiliation and research purpose. Access requests will be reviewed and granted after verification is complete for non-commercial academic research.
Log in or Sign Up to review the conditions and access this dataset content.
MGL Amsterdam Cross-View Geo-Localization Dataset
License & Dataset Verification in Progress: Dataset packaging, coordinate alignment, and third-party licensing terms are currently undergoing active verification. Access requests submitted through the gated form will be reviewed, and download access will be granted after verification is complete for authorized non-commercial academic research.
Dataset package for 3-DoF cross-view visual geo-localization (lateral translation, longitudinal translation, continuous yaw angle) in Amsterdam, pairing perspective ground images from Mapillary with aligned satellite orthophotos from Google Static Maps.
Dataset Structure
mgl-amsterdam-crossview/
├── metadata/
│ ├── splits_amsterdam_resolved.json # Sample-level attributes (27,159 train, 9,103 val)
│ ├── splits_amsterdam_exact.json # List of active perspective view IDs
│ ├── splits_amsterdam.json # Mapillary sequence partition index
│ └── satellite_pair.pkl # Ground view to satellite tile mapping
├── data/
│ ├── amsterdam_satmap_seq_full.tar.gz # Satellite crops (640x640, Zoom 18, Scale 2)
│ ├── satmap_seq_full/ # Extracted satellite crops
│ └── images/ # Ground perspective images (downloaded via script)
└── scripts/
├── mgl_amsterdam_dataset.py # PyTorch Dataset loader (auto-extracts satellite archive)
├── download_mapillary_mgl.py # Mapillary Graph API downloader
├── download_satellite_tiles.py # Google Static Maps tile fetcher
├── prepare_amsterdam_dump.py # Satellite pair mapping generator
├── convert_splits.py # Sequence-to-view split converter
├── export_resolved_splits.py # Dataloader split exporter
└── verify_dataset.py # Integrity and split disjointness checks
Data Access and Setup
1. Satellite Imagery
- Option A (Hugging Face Gated Archive): Request access on this repository and place
amsterdam_satmap_seq_full.tar.gzintodata/. The PyTorch loader automatically extracts it on first run. - Option B (Direct Google Maps API): Fetch tiles directly using
scripts/download_satellite_tiles.pywithGOOGLE_MAPS_API_KEY.
2. Ground Images and OpenSfM Poses
Download ground images using the Mapillary Graph API v4:
export MAPILLARY_ACCESS_TOKEN="your_token"
python scripts/download_mapillary_mgl.py
To also fetch official OpenSfM 6-DoF camera poses into metadata/dump.json:
python scripts/download_mapillary_mgl.py --fetch_sfm_poses
3. Verification
Verify dataset partition counts and metadata integrity:
python scripts/verify_dataset.py --root .
Dataloader Usage
from torch.utils.data import DataLoader
from scripts.mgl_amsterdam_dataset import MGLAmsterdamDataset
train_dataset = MGLAmsterdamDataset(
root_dir=".",
split="train",
rectify_pitch_roll=True, # Level horizon (pitch=0, roll=0)
auto_extract=True
)
val_dataset = MGLAmsterdamDataset(root_dir=".", split="val", rectify_pitch_roll=True)
loader = DataLoader(train_dataset, batch_size=16, shuffle=True, num_workers=4)
for batch in loader:
# batch["ground_image"]: [B, 3, 512, 512]
# batch["satellite_image"]: [B, 3, 512, 512]
# batch["gps_latlong"]: [B, 2]
# batch["yaw_deg"]: [B]
break
Dataset Specifications
- Active Splits: 27,159 train views, 9,103 validation views (36,262 total views; 50% left/right perspective views filtered).
- Satellite Resolution: Zoom 18, Scale 2, 640x640 pixels (~0.1804 m/pixel at Latitude 52.368°).
- Satellite Capture Window: November 6–9, 2024.
- Ground Image Window: January 28, 2016 to September 29, 2020.
- Coordinate Reference: WGS84 coordinates mapped to local metric tangent plane; continuous heading angle $\theta \in [-180^\circ, 180^\circ]$ relative to True North.
Attributions and Licenses
- OrienterNet: Dataloader structure, attitude rectification (
rectify_image), and preprocessing tools are adapted from OrienterNet (Paul-Edouard Sarlin et al., Meta AI / ETH Zurich, CVPR 2023), licensed under Apache 2.0. - CVGL Utilities: Tile retrieval tools adapted from spatial-aware cross-view geo-localization pipelines (Yujiao Shi et al., CVPR 2019 / TPAMI 2020).
- Images: Ground perspective images are subject to Mapillary Terms of Use. Satellite orthophotos are copyright of Google and respective imagery providers. This dataset is provided strictly for non-commercial academic research.
- Downloads last month
- 4