Vesuvius challenge grand prize submission (runner-up)

Winning Solution for Vesuvius Challenge First Title. Custom MiniUNETR 3D architecture for ink detection in carbonized scrolls.

Unveiling the secrets of an ancient library buried by Mount Vesuvius, this Kaggle competition, supported by the Vesuvius Challenge organization, tasked participants with detecting ink from 3D X-ray scans of charred scrolls preserved in a Roman villa in Herculaneum.

Extended implementation of the Vesuvius Challenge 2023 Grand Prize winner. This project features extra functionality compared to the original ink detection script, giving users fine control over the layers they perform inference on.

3 Fragments 1 Hour >94% F1

Agent-assisted open-source toolkit for detecting and reviewing suspicious regions in reconstructed papyrus surfaces

Vesuvius Challenge - Surface Detection (Kaggle) · 🥈 Silver Medal · nnUNetv2 + MONAI 3D-SegM baselines

Vesuvius challenge topological and geometric analysis of scrolls and fragments

Machine learning pipeline for the Vesuvius Challenge an open research effort to read carbonised Herculaneum scrolls without unrolling them. Uses PyTorch convolutional models on 3D X-ray micro-CT data to segment ink signatures from charred papyrus, contributing to the recovery of ancient texts

3D surface segmentation of ancient Herculaneum papyrus scrolls from micro-CT scans. Dual-pipeline architecture: V1 (CE+Dice, clDice) and V2 (Focal Loss, Deep Supervision, Skeleton Recall). Built for the Vesuvius Challenge $200K Kaggle competition. CC BY-NC 4.0.

ONNX C/C++ Runtime to detect ink in Vesuvius scrolls

Direct surface detection from raw CT sinogram (projection) data for Herculaneum papyrus scrolls, bypassing the traditional reconstruct-then-segment pipeline. Explores whether discriminative signal lost during filtered back-projection can improve detection of tightly compressed, low-contrast papyrus layers. Built for the Vesuvius Challenge.

Optimize segmentation for Vesuvius Project

Certified topology auditing and transverse-clean tifxyz outputs for a five-scroll corpus. Report-only by default; finds where a traced surface passes through itself.

Calibration benchmark for winding constraint generators, sealed criteria, human-annotation ground truth

Trace-quality reports for Vesuvius Challenge scroll segments — flags likely segmentation errors from an ink-detection prediction (CPU-only).

SUBIT-INK replaces the standard binary ink classifier with a four-valued Belnap bilattice, enabling the model to express epistemic states beyond a simple yes/no decision.

Registered human ground-truth ink evaluation for the open Vesuvius Challenge SOTA scroll data

Label-free local layer-separability diagnostics for Vesuvius Challenge CT volumes