gaetanbahl/Global-Scale-Road-Dataset

Global-Scale Dataset for Road Graph Extraction

Disclaimer: This is an unofficial re-upload of the Global-Scale dataset for researchers who do not have easy access to Baidu. All credit goes to the original authors. If the authors have any concerns with this mirror, please contact me and I will remove it immediately.

Original Source

• Paper: Towards Satellite Image Road Graph Extraction: A Global-Scale Dataset and A Novel Method (CVPR 2025)
• Official Repository: https://github.com/earth-insights/samroadplus
• Original Dataset Location: Baidu Pan

Citation

If you use this dataset, please cite the original authors:

@inproceedings{yin2025towards,
  title={Towards Satellite Image Road Graph Extraction: A Global-Scale Dataset and A Novel Method},
  author={Yin, Pan and Li, Kaiyu and Cao, Xiangyong and Yao, Jing and Liu, Lei and Bai, Xueru and Zhou, Feng and Meng, Deyu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2025}
}

Authors

• Pan Yin^1,2*, Kaiyu Li^4*, Xiangyong Cao^1,2 (corresponding), Jing Yao⁶, Lei Liu⁷, Xueru Bai⁷, Feng Zhou⁷, Deyu Meng^2,3,5

¹School of Computer Science and Technology, Xi'an Jiaotong University ²Ministry of Education Key Laboratory of Intelligent Networks and Network Security, Xi'an Jiaotong University ³School of Mathematics and Statistics, Xi'an Jiaotong University ⁴School of Software Engineering, Xi'an Jiaotong University ⁵Pengcheng Laboratory ⁶Chinese Academy of Sciences ⁷Xidian University

Dataset Description

The Global-Scale dataset is a comprehensive road graph extraction dataset designed for training and evaluating road network extraction models from satellite imagery. It is approximately 20x larger than the largest existing public road extraction dataset and spans over 13,800 km² globally.

Key Features

Property	Value
Total Images	3,468
Image Size	2048 × 2048 pixels
Spatial Resolution (GSD)	1.0 m/pixel
Coverage	~13,800 km²
Geographic Scope	Global (all continents except Antarctica)
Region Types	Urban, Rural, Mountainous
Label Type	Graph labels (nodes and edges)
Year	2024

Dataset Splits

Split	Images	Description
Train	2,375	Training set
Validation	339	Validation set
Test (In-Domain)	624	Test images from regions represented in training
Test (Out-of-Domain)	130	Test images from unseen regions (Hong Kong, Shenzhen, Lucerne)

Data Sources

• Satellite Images: Google Earth / Google Static Map API
• Road Graph Labels: OpenStreetMap (with manual verification for annotation completeness)

Comparison with Other Datasets

Dataset	Graph Label	Size	Train	Val	Test-ID	Test-OOD	GSD	Region	Region Type
Massachusetts	No	1,500²	1,108	14	49	-	1.0	Massachusetts	U, R
DeepGlobe	No	1,024²	6,226	243	1,101	-	0.5	Thailand, Indonesia, India	U, R
SpaceNet	Yes	400²	2,167	-	567	-	0.3	Paris, Las Vegas, Shanghai	U
City-Scale	Yes	2,048²	144	9	27	-	1.0	20 cities in U.S.	U
Global-Scale	Yes	2,048²	2,375	339	624	130	1.0	Global	U, R, M

U = Urban, R = Rural, M = Mountainous

Dataset Structure

HuggingFace Folder Organization

Due to HuggingFace's limitation of 1,000 files per folder, files have been organized into numbered subdirectories. Each subdirectory contains 100 tiles (500 files), grouped by tile number:

• Folder 1: tiles 0-99
• Folder 2: tiles 100-199
• Folder 3: tiles 200-299
• ... and so on

Global-Scale/
├── train/                    # 3,338 tiles total (train + val + in-domain test)
│   ├── 1/                    # tiles 0-99 (500 files)
│   ├── 2/                    # tiles 100-199 (500 files)
│   ├── ...
│   └── 34/                   # tiles 3300-3337 (190 files)
├── val/                      # 339 tiles
│   ├── 1/                    # tiles 0-99
│   ├── 2/                    # tiles 100-199
│   ├── 3/                    # tiles 200-299
│   └── 4/                    # tiles 300-338
├── in-domain-test/           # 624 tiles
│   ├── 1/                    # tiles 0-99
│   ├── ...
│   └── 7/                    # tiles 600-623
└── out_of_domain/            # 130 tiles (Hong Kong, Shenzhen, Lucerne)
    ├── 1/                    # tiles 0-99
    └── 2/                    # tiles 100-129

Files Per Tile

Each tile consists of 5 files:

• region_X_sat.png - Satellite image (2048×2048)
• region_X_gt.png - Ground truth mask
• region_X_graph_gt.pickle - Graph ground truth
• region_X_refine_gt_graph.p - Refined graph ground truth
• region_X_refine_gt_graph_samplepoints.json - Sample points for the refined graph

Original Structure Note

In the original dataset, the train folder includes training, validation, and in-domain test data together for convenience during training. Data partitioning rules are defined in dataset.py (see the official repository).

Flattening the Structure

If you need the original flat structure, you can flatten each directory with:

# Linux/Mac
cd train && mv */* . && rmdir */

# PowerShell (Windows)
Get-ChildItem -Directory | ForEach-Object { Move-Item "$($_.Name)/*" . }; Get-ChildItem -Directory | Remove-Item

Usage

For data preparation and training instructions, please refer to the official SAM-Road++ repository.

License

Please refer to the original repository and paper for licensing information. This re-upload is provided solely for research accessibility and does not claim any ownership or modification rights over the original dataset.

Acknowledgments

This dataset was created by the authors listed above. The collection methodology involved:

1. Manual selection of longitude/latitude coordinates for various road types using Google Earth
2. Satellite image collection from Google Static Map API
3. Road graph data collection from OpenStreetMap
4. Manual verification to ensure annotation completeness

Contact

• For questions about the dataset or methodology, please contact the original authors via the official repository
• For issues with this HuggingFace mirror, please open an issue on this repository