mchelali/forbin_dataset

Forbin Dataset: A collection of historical photographs with archival metadata

This repository hosts the Forbin Dataset, a large-scale collection of historical photographs taken or collected by Victor Forbin (1868–1947).

This HuggingFace dataset version provides:

• COCO-style annotations (segmentation polygons)
• Archival metadata (Box ID, description, notes, dates when available)
• A lightweight explorer interface (HTML/JS) to preview images and annotations: https://mchelali.github.io/forbin_dataset/

📜 Dataset Description

The Forbin Dataset contains digitized historical photographs from the personal archives of Victor Forbin, a French explorer, photographer, and writer.
Images are accompanied by rich metadata and manually extracted segmentation polygons suitable for:

• Computer Vision
• Document Analysis
• Cultural Heritage Studies
• Machine Learning Research

The sample included here is intended for illustration and early experimentation only.
The upcoming full release will contain tens of thousands of images with complete metadata and annotations.

🛠️ Data Access and Usage Instructions

Given the size of the image archives, the dataset must be loaded in a two-step process: Local Download followed by Indexing.

1. Downloading the Raw Data Files (Images and Annotations) ⬇️

The dataset is distributed as WebDataset archives (.tar) and separate JSON annotation files. You must download these files locally before starting the training process.

File	Content	Note
forbin_all.json	All Image IDs, metadata, and annotations (for annotated images).	Used for full dataset indexing.
forbin_annotated.json	Only images that have associated annotations (simplified index).	Useful for training on annotation tasks.
data/*.tar	WebDataset archives containing all raw images.	Large files.

Mode A: Via the Hugging Face Command Line Interface (CLI)

This is the fastest method for users familiar with the terminal.

# Requires installation: pip install huggingface_hub
hf download  mchelali/forbin_dataset  --repo-type dataset --local-dir ./forbin_data_local 

Mode B: Via Python (Recommended for Resumable Downloads)

This reliable method uses the official Python API, which automatically handles resuming the download process if interrupted.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="mchelali/forbin_dataset",
    repo_type="dataset",
    local_dir="./forbin_data_local" # Your chosen destination folder
)

Web Download Interface (For SHS Researchers):

For users less familiar with the command line, we provide a dedicated web interface to download the individual .tar archives one by one:

➡️ Web Download Interface: https://mchelali.github.io/forbin_dataset/download.html

---

2. Indexing and Annotation Usage 📚

Once the *.json and *.tar files are downloaded locally, you can build your own data loading pipeline.

Annotation Format:

All annotations (including textual metadata, bounding boxes, and segmentation polygons) are provided in the standard COCO (Common Objects in Context) format. This ensures compatibility with existing computer vision tools and libraries like PyTorch, TensorFlow, and pycocotools.

The JSON file acts as your Manifest (Index Table). It links the image ID (via image_id) to the image's location within the .tar archives (via the file_names field in the images section).

To use the dataset:

1. Load the JSON file (forbin_all.json or forbin_annotated.json) into your program.
2. Use the Python tarfile (or webdataset) library to open the corresponding .tar archive and load the image bytes based on the path provided in the file_names field.
3. Apply the COCO annotations (found in the annotations section of the JSON) to the loaded image.

🔖 License

This sample dataset is released under the following license:

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
➡️ https://creativecommons.org/licenses/by-nc/4.0/

This means:

• ✔ You must provide attribution
• ✔ You may share and adapt the material
• ❌ You may not use it for commercial purposes

📚 Citation

If you use this dataset or the sample in academic work, please cite the forthcoming data paper:

[Under review] 
Chelali M., Gosselet S. K., Cloppet F., Kurtz C., Bloch I. and Foliard D.,
The Forbin Dataset: A collection of historical photographs with archival metadata, 2025.

🤝 Acknowledgment of Authors

This dataset originates from the personal archives of Victor Forbin, digitized and curated by the High Vision Project – Archives & Vision Initiative.
All annotation and data processing work was performed by the project contributors.

This work is supported by the French National Research Agency under the ANR-24-CE38-4079 project