May Progress Prizes and Updates to Tooling
Working with scroll data can be challenging. The data is quite large (often over a terabyte for a single scroll), the tooling for segmentation and processing can be difficult to install or confusing to run. The pipeline from raw data to readable segmentations can (and often is) quite daunting, especially to those less familiar with software engineering and general scientific python usage.
This month’s prizes coincided with a focus from our team to make this pipeline easier.
1 x $5,000
Philip Allgaier / @spacegaier made some significant improvements to VC3Ds overall stability and ease of use. The largest improvement is the return of prebuilt docker packages. Anyone who has built volume-cartographer from source can attest to the difficulties that can arise, particularly from Qt. No longer!
Philip also addressed a number of other issues, many from our issues list on the volume-cartographer repository:
Prevented memory leaks in multiple VC3D command line tools (such as
vc_grow_seg_from_segments) as well as in VC3D itself (GH issue-16)Added Github build / CI actions that create Docker containers and upload to ghrc.io (GH issue-10)
Ensured compatibility with Ubuntu 25.04 & QT 6.8 (resolved surface tree widget promotion path problem in QtDesigner)
Improve console output (added missing blanks, clearer messages)
A full list of changes Philip made is here.
2 x $1,000
Marcel Roth / @_mvrcii implemented a rapid pipeline for masking scrolls from their case and background. Masking has been explored in the past and @james_darby implemented a SAM based masking model. SAM however comes with some larger gpu requirements and at times can be difficult to run.
Marcel used a model that was initially trained as a proof of concept by a member of the challenge team on a very small sample size, and while the model itself definitely needs more training, it’s clear that this method should work quite well. It’s extremely fast (all of scroll 4 in only 2 minutes!) and the requirements are low enough that it can be ran on just about any GPU accelerated device.
Yao Hsao / @.yaohsiao and Dalufish / @dalufish continued their improvements on a fork of neuroglancer called neuroglancer-mini, implementing a local-first design philosophy by caching data locally and removing reliance on the chromium file system api, and creating a more detailed readme.
More VC3D Updates!
In addition to the work of Philip Allgaier / @spacegaier , the Vesuvius Challenge team has been hard at work with improvements to the general workflow within VC3D:
Most tools are now runnable via a right-click context menu — tracing, rendering, obj conversion, even deletion!
Traces now output to a ‘traces’ directory, within the volpkg, and can be viewed by clicking the dropdown under the segmentation directory and the volume views are filtered for these, along with the list of segmentations.
A number of additional seeding methods have been added, with the goal of reducing computation on areas outside of the target area for tracing (like the scroll casing). These are all runnable via a widget in the GUI
Seeding Widget Radial seeding based on intersections over a distance transform
Drawn Path based seeding which also uses intersections over a distance transform
Drawn path and Radial based seeding
Freehand annotation / drawing is supported in the volume views, and works in full 3D (the saving is not yet implemented, but this can be a handy way to visualize how patches or traces overlap or intersect)
Additional filters and tags, and the ability to use multiple at once
Reviewed tag / filter
Revisit tag / filter
Partial Review tag / filter (ones that overlap with reviewed)
Defective Filter
Expansion Filter (filters out seeds created via expansion mode)
Extended Filters
VC3D is becoming much easier to work with, and many of these changes have made for large productivity improvements. More changes to VC3D and vesuvius python coming your way soon!
Don’t forget to check out our prizes page for up-to-date information on our current open prizes, and join the discord or check out our get-started page to compete for these!