Vesuvius Challenge Progress Prizes: December Edition

Ending 2024 with more great advancements

Next up in the 2024 year-end prize series, we are honoring the community advancements submitted for December Progress Prizes!

There’s a heavy focus right now on dialing up the data quality: more and better ink labels, more and better segmentation inputs, refining existing versions of each, better model architectures for segmentation tasks, and so on. This inexorable march forward is illuminating an increasingly clear path towards automated virtual unwrapping. Much remains to do, but the obstacles are going down in size and number all the time…

[potential spoilers] Some submissions for December prizes overlapped heavily with First Automated Segmentation Prize submissions, so, you know. Maybe you’ll hear more about those soon!

For now: these fantastic December contributions!

3 x $2,500 (Sestertius)

• Tim Skinner / @Tim S, Lachlan Parker and John Skinner trained a volumetric segmentation model using the team/community generated instance annotation cubes that is capable of some impressive inference outputs. We’re excited to see this addition to our growing family of segmentation models!

  

  16 adjacent volumetric patch inferences (manually selected/stitched together out of inference cubes).

• Nicola Bodill / @polytrope has contributed a body of ink labels across our scroll datasets! These have been used successfully by other community members for training ink detection models- check out in particular the PHerc. 172 (Scroll 5) labels. This tedious annotation is crucial for improving ink detection!

  

  Example label from inside PHerc. 172. Preliminary annotations, not meant to be perfect - reach out to @polytrope on Discord if you have refinements!

• James Darby / @james darby developed a fork of Khartes (documentation w/ instructional videos) that allows one to interactively refine existing segmentation meshes. This can be used to fix up segmentations as a final product, or to improve them before using them as machine learning inputs. For both use cases, we are excited to see interactive segmentation improvement tooling!

  

  Interactively refining an existing mesh segmentation.

6 x $1,000 (Papyrus)

• Johannes Rudolph / @jrudolph developed vesuvius-render, a self-contained CPU-based alternative to vc_render and scroll-renderer.

  

  Rendering a segment into layers (with nice progress bars, too!).

• Yao Hsiao / @Yao Hsiao and Dalufishe / @Dalufish continued work on their electric potential field-based scroll flattening approach, also adding a couple of command line tools for using it!

  

  Using simulated electric potential fields to locally flatten scroll regions!

• Chuck / @khartes_chuck added overlay capabilities to Khartes! This is a great way to visualize various data formats (largely segmentation volumes) overlaid on the raw scroll data - check it out using your datasets!

  

  Scroll data overlaid by detected fibers (blue horizontal, red vertical) and detected ink (yellow) - all in one view!

• Hendrik Schilling / @waldkauz significantly improved inference times for ink detection, and added some command line settings for tuning the output image quality!

  

  Different output quality settings for ink detection images.

• Forrest McDonald / @verditelabs shared a set of segmentation techniques, data, and visualizations, based largely on supervoxels (a 3D extension of superpixels) and on merging them together into “chords” that span the entire scroll height.

  

  Voxel rendering of a scroll chunk.

• William Stevens / @Will Stevens added to a growing set of segmentation tools, this time with particle-based surface flattening and interpolation, as well as simulated annealing for patch assembly. As always with Will’s work, check out the excellent write-up!

  

  Filling holes while preserving geometry in a test segmentation patch.

2024 has been an incredible year for communal progress on this great technical challenge - and we haven’t covered it all quite yet. Don’t touch that dial!