ScrollPrize/villa

What it solves

This project provides a comprehensive suite of tools for the Vesuvius Challenge, a competition to read ancient Herculaneum scrolls using machine learning and computer vision. It addresses the challenges of virtually unwrapping scrolls (segmentation) and detecting ink on carbonized papyrus from CT scans.

How it works

The repository is a collection of libraries and pipelines for different stages of the pipeline:

• Access and Training: The vesuvius library handles CT scan access and ML model training/inference for surface prediction.
• Ink Detection: The ink-detection toolset uses models based on the Grand Prize winner of 2023 to identify ink on the scrolls.
• Volumetric Learning: dinovol implements a self-supervised 3D DINO for learning volumetric representations.
• Unwrapping (Segmentation): Multiple pipelines are provided, including VC3D (a semi-automatic surface tracer), lasagna (a PyTorch-based optimization framework for surface meshes), and spiral fitting (an automatic pipeline for fitting a global spiral to a scroll).

Who it’s for

Researchers, computer vision experts, and researchers of ancient texts who are using CT scans to read carbonized scrolls without physically opening them.

Highlights

• Diverse Unwrapping Methods: Offers multiple approaches to fiber tracing and surface extraction from 3D volumes.
• Grand Prize Winning Tech: Includes tools for ink detection based on winning competition entries.
• 3D Representation Learning: Uses self-supervised learning to identify patterns in CT scan data.
• Integrated Ecosystem: Combines data processing, cloud infrastructure management, and model checkpoints via Hugging Face.