A private archive in Pakistan, the Ibteda Digital Library, has repurposed a decade of Photoshop work into training data for AI. The library, run by three individuals, digitized over 1,700 rare Urdu books using a self-made setup with a Nikon DSLR and basic lighting. Their process involved capturing images of each two-page spread, including imperfections like shadows and stains, which were then meticulously edited in Photoshop. This extensive manual effort, totaling over 10,000 hours, resulted in 575,729 processed pages, with the camera exceeding its shutter durability by nearly six times. Faced with funding issues and personal circumstances, the library ceased operations, leaving over half a million pages unprocessed, but plans to release the code and models used for this unique dataset. AI
IMPACT This project demonstrates a novel method for creating labeled training data from manual editing processes, potentially inspiring new approaches in data curation for AI.
RANK_REASON This item describes a unique, non-commercial project using existing tools for data creation, rather than a new release, significant industry event, or academic research.
Read on Mastodon — mastodon.social →
- Garibaldi
- Ibteda Digital Library
- Internet Archive
- Nikon D5300
- Photoshop
- scale-invariant feature transform
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
