Open Problems: Why Reading Every Herculaneum Scroll Is Still a Challenge
8 hours ago
- Scanning Herculaneum scrolls requires synchrotron X-ray micro-CT with careful tuning of voxel size, X-ray energy, and sample-detector distance to balance resolution and phase contrast.
- Unwrapping involves surface localization, mesh tracing, and flattening; tools like GrowPatch, lasagna, and spiral fitting are used, but human correction is still needed for errors like holes, mergers, and sheet switches.
- Ink recovery is challenging due to carbon ink on carbonized papyrus; models trained on fragments use pseudo-labeling and self-distillation, but generalization across scrolls remains a bottleneck.
- Label quality is a major bottleneck—approximate surface and fiber labels limit model accuracy, so better annotation methods like label snapping and active learning are needed.
- Data scale is an infrastructure challenge; cloud-native formats like OME-Zarr enable streaming of terabyte-scale volumes, facilitating collaborative research.
- Future directions include self-supervised learning (e.g., 3D DINO), direct 3D ink segmentation, and improving diagnostics to distinguish 'no ink' from 'ink not yet recovered'.
- The Vesuvius Challenge aims to make the pipeline fully automatic and reliable for all scrolls by June 2027, with six key areas for community contribution.