Last week we announced $35,000 in Segmentation Tools Prizes, and since then the community has been hard at work to build better segmentation software (but there is still plenty of room for you to join in, so join us on Discord!).
So far we’ve published a couple of segments in Scroll 2, such as this set of 3 adjacent segments:
I added the capital “rho” for scale. The segments are available on the data server, in the /full-scrolls/Scroll2.volpkg/paths/
directory. They can be opened in Volume Cartographer (see Tutorial 3: Segmentation and Flattening). Perhaps you can be the first to detect some actual writing in these segments!
If you want to compete for the Segmentation Tools Prizes, and need inspiration, try making our segmenters more efficient! We put together a list of feature requests that we’d be grateful to see addressed.
We are also still looking for contractors to create more segments! Please message us on Discord for more information.
Community news
There have been lots of discussions on how to evolve Volume Cartographer and VolumeAnnotate to make segmenters more efficient. We seem to be moving to a world where Volume Cartographer is the high-performance algorithm library, and VolumeAnnotate its GUI frontend, though things are moving fast so anything can change!
James Darby used Meta AI’s Segment Anything model to remove non-scroll data from the 3d scroll volumes. These resulting volume is about 2x smaller, and will be available on the data server soon.
Santiago Pelufo created another data transformation for the scrolls: chopping up the 3D volume into smaller 3D “grid cells”, which allows for more efficient loading when only looking at one part of a scroll. These are now available on the data server, e.g. at
/full-scrolls/Scroll1.volpkg/volume_grids/
.Brett Olsen discovered super-efficient loading of tiff files in Python, and he has been working on a different segmentation algorithm, though it is still mysterious..
Pavel Hanchar has been experimenting with statistical methods which suggest different Gaussian distributions for ink vs no-ink.
Andrii Kapatsyn posted great tips for data loading in Kaggle.
Moshe Levy made big strides with VolumeAnnotate, including partial Volume Cartographer compatibility, faster data loading, and 3D preview. James Darby contributed “jump to slice” functionality.
Luke Farritor has made Scroll Viewer work with fragments.
We updated the Kaggle ink labels with some small corrections. Thanks “Mecndo” and “JavaZero” on Kaggle for reporting!
Jessica Hamel created Stack Scanner for running various algorithms on .tif files.