The hard parts are getting clearer

May progress update

Over the past few months, the Vesuvius Challenge team and community have continued to push the boundaries of what is possible.

Our papyrology team lead, Prof. Federica Nicolardi, is organizing a conference in Naples from June 22 to 26 on recent advances in Herculaneum papyrology. The conference is supported by the Vesuvius Challenge, which will also hold a press conference on recent updates on June 25. Click here to read the program.

A livestream link will be available.

Technical updates

What still isn’t working as well as we hoped

We are close to saturating the performance of models trained on the approximate surface labels our annotators can produce.

The training recipes from the recent Kaggle competition, together with the new high quality data, made it possible to segment and unwrap many regions of the scrolls that previously seemed impossible.

Piercing through compressed regions. Difficult but doable.

However, our representations still struggle in some of the most densely packed regions, which are not rare, and also in regions characterized by either a twisted or broken shape. These regions still defeat fully (and partial) automated approaches.

Very hard spot. Surface predictions disappear and barely make sense.

A hard spot requiring a lot of effort since the surface predictions disappear or form blobs.

Another example of a hard region to segment.

Since humans can, in principle and with effort, trace sheets throughout the scan, we believe models could do the same if given the right supervision. However, producing this supervision is not easy and manual annotation does not scale.

This creates a bottleneck for both segmentation via surface detection models and all the derived approaches, for instance Neural Tracing which soon drifts off the intended surface (and we also realized that inference is very slow).

The future likely lies in training self-supervised methods to extract relevant features without (or with minimal) annotation. The team in the past months trained a 3D version of DINO, and will keep working on this in the next weeks.

Lasagna and Spiral Fit

Hendrik added several new features to the Lasagna workbench for VC3D:

1. Semi-automated fiber tracing. Where the latent representation holds, an annotator can trace a fiber at the speed of 2 cm / minute.

2. Inpainting. The Lasagna representation can now more efficiently bridge the gaps where the original surface predictions are noisy.

   

   Region of mesh to inpaint

How the inpainted mesh looks in 3D

Paul and Sean returned to the diffeomorphic spiral fit, recent algorithmic changes and the extended capability to let it conform to different kinds of annotations (patches, skeletons, points) made it much more flexible and capable of adapting to more complicated geometries.

It still does not conform well where annotations—whether automatic or manual—are missing.

May 2026 Progress Prizes

The team received many submissions this month! After careful evaluation, we decided to award:

• $10,000 to Ben Kyles, former Segmentation Team Lead, for Scroll Fiesta, an automatic mesher of surface predictions that attempts to fix topological mistakes. The work is still at an early stage, and the team would like to encourage this effort and see it very soon integrated into the current VC3D toolkit in a faster and more robust form.

  

  Output image of ScrollFiesta, mesh created by welding fixed connected components across adjacent regions.

• $2,000 to Paulo Sergio Camillo for introducing three 3D augmentations to train ML models on the CT scans, now merged into the vesuvius library.

  1. Scroll Decohesion

  2. Realistic Warp

  3. Squeeze

     

     Example showing how the Warp transform is more realistic than the Elastic transform.

Huge congratulations to the winners!

Feel like contributing? Join our Discord channel and find out how!