HeritageNet: A Dataset of Nepali Heritage Sites

A collection of photographs of 26 Nepali heritage sites — temples, stupas, and monasteries — captured by the HeritageNet team for image-classification research.

Nepali cultural heritage is severely underrepresented in existing computer-vision datasets. Global models recognise landmarks like the Eiffel Tower or Taj Mahal, yet have effectively never "seen" Nepal's local heritage. This dataset is a small step toward closing that gap — putting Nepali heritage on the digital map and enabling on-device recognition of these sites.

Dataset summary

• ~1,044 images across 26 classes (one folder per heritage site).
• Photographed by the HeritageNet team.
• Intended for image classification (identify which heritage site a photo shows).
• Released under CC BY 4.0 — free to use, share, and adapt with attribution.

Structure

Images are organised in the standard image-folder format — one subfolder per class:

Ashok_stupa_Chaitya/
Badrinath_Temple/
Balkumari_Temple/
... (26 site folders) ...

Each subfolder contains JPEG photographs of that site. The folder name is the class label. This layout loads directly with common tools (e.g. torchvision ImageFolder, or the Hugging Face datasets imagefolder loader).

How it was collected

The images were photographed directly by the team at the heritage sites, using consumer phone cameras, as part of the HeritageNet project. Capturing our own photographs — rather than scraping the web — is central to the project: it produces a dataset we can openly and lawfully release, and it reflects the "citizen science" vision of ordinary people helping to document heritage with the phones they already carry.

Intended uses

• Training and evaluating heritage-site image classifiers.
• Educational and research use in computer vision and cultural-heritage documentation.
• A base others can extend with additional sites and images.

A companion lightweight classifier (MobileNetV3-small, transfer learning) trained on this data reaches ~86.6% test accuracy / ~85.8% macro-F1 across the 26 classes, and runs on-device in the browser.

Known limitations

• Class imbalance / thin classes. Some sites have relatively few images. In particular, the Surya_Binayak_Temple class currently contains only 22 team-photographed images. Surya Binayak will be expanded with additional team photographs in a future version.
• Placeholder class names. A few classes retain provisional names (e.g. Heritage_1, Heritage_2, Heritage_3) and will be renamed to their proper site names in a future release.
• Regional scope. The dataset covers 26 specific sites and is not exhaustive of Nepali heritage. It is intended to grow over time.
• Photographic variation. Images vary in lighting, angle, and background; some may include incidental passers-by.

Roadmap

This is an initial release. Planned future work: expand the number of sites, add more images per class (especially the thin ones), rename placeholder classes, and open the dataset to community photo contributions.

License

This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You are free to share and adapt the material for any purpose, even commercially, provided you give appropriate credit. License: https://creativecommons.org/licenses/by/4.0/

Citation / Attribution

If you use this dataset, please credit:

HeritageNet: A Dataset of Nepali Heritage Sites (2026). Collected by Ashutosh Chapagai, Rasim Mahato, Diva Awasthi, Kriti Puri, Suyog Adhikari, and Pratik Sharma. Mentors: Sijan Shrestha and Hari Om Sah. Licensed under CC BY 4.0.

Acknowledgements

Created as a research project to help document and preserve Nepali cultural heritage. Mentored by Sijan Shrestha and Hari Om Sah.