4th fragment released; prize deadlines coming up; ink crackle hypothesis
Lots of progress this summer
Upcoming prize deadlines
August 28th is the deadline for submitting for the $12,000 Ink Detection Followup Prize, as well as for the $50,000 Segmentation Tooling 2 Prize.
As a reminder, here are some things we would love to see for the Ink Detection Followup Prize:
Simply running the open source models on the many segments we have available on the data server (over 600cm^2!). Negative results and analysis posts also qualify!!!
Retraining open source models on downscaled volumes, to simulate 8 micron resolution instead of 4 micron resolution. We’re extremely curious how much this difference in resolution matters for model performance, so an analysis of this would definitely qualify!
Adapting or fine-tuning open source models for the full scrolls, or tooling to generate ground truth annotations to do this (e.g. based on the new crackle hypothesis, see below).
More ideas listed here!
For the Segmentation Tooling 2 Prize, we’re looking for tooling to help create large or hard segments. Those two are very related, since the larger a segment, the more likely that you’ll encounter hard regions (e.g. with papyrus layers touching or even fused). Scroll 2 has been especially hard to crack!
Finally, a reminder that the $50,000 First Letters Prize is still ongoing.
Releasing the 4th fragment
The secret 4th fragment that was used for verification in the Ink Detection Prize has been released on the data server! This can now be used to fully reproduce the results of that prize, as well as train new models on.
We’re also revealing what 4 fragments are. They are: Fragment 1: P.Herc.Paris. 2 fr. 47; Fragment 2: P.Herc.Paris. 2 fr. 143; Fragment 3: P.Herc.Paris. 1 fr. 34; Fragment 4: P.Herc.Paris. 1 fr. 39. The data paper has also been updated with this information, and contains more details.
Ink crackle hypothesis
Casey Handmer wrote a great blog post about his discovery of “crackle” patterns within the full scrolls, which might correspond to presence of ink.
The community is hotly debating whether this holds up (e.g. by looking at the fragments for which we know where the ink is), finding more areas in the scrolls that show these patterns, and trying to apply machine learning to find these areas. It’s an extremely promising line of inquiry.
Community news
Stephen Parsons from the University of Kentucky (and from our Review Team!) has published his PhD dissertation, with countless gems.
Santiago Pelufo has continued work on working with segments in Blender.
Yao has visualized our segments so far.
emel_ryan applied their winning Kaggle model to the Monster Segment, but sadly no dice!
WayneWayneHello has continued CT-scanning his own carbonized papyrus! With the scans being available on our data server.
Brett Olsen has been looking at the crackle patterns using cross-sections.
Thank you. Keep up the great work!