Ink Detection Followup Prize Winners!
We ran the Ink Detection Followup Prize to encourage people to take the results from the larger Ink Detection prize on Kaggle and apply them to the scrolls. This new prize finished last week, and we have selected our winners.
$3,000 — Ryan Chesler
Ryan wrote up an extensive analysis of applying their winning Kaggle solution to the scrolls. It looks like it predicts a few letters, though it’s too hard to tell.
He also looked at the effect of resolution on the model’s performance on the original fragments, suggesting that the difference in resolution between the fragments and the scrolls is significant:
$3,000 — Luke Farritor
Luke fine-tuned one of the Kaggle models on Casey’s ink crackles, in the hope to find more interesting areas that way. This particular model didn’t seem to reveal new letters, but it’s exactly the sort of approach we were looking for.
$1,500 — Julian Schilliger
Julian made a tool for annotating ink crackles, which helps in fine-tuning models.
$1,500 — Team OverthINKingSegmenter
Team OverthINKingSegmenter did a great analysis on the effect of resolution on models, very much in line with Ryan’s findings.
$1,500 — Youssef Nader
Youssef suggests an interesting approach for pre-training on the scrolls, and then fine-tuning on the fragments.
$1,500 — Yao and Dalufishe
Yao and Dalufishe have been working on a data exploration tool for looking at model predictions, annotations, and raw fragment and scroll data in one interactive canvas.
Discord has been buzzing with activity lately. Some highlights:
Luke Farritor has been improving his models, and potentially finding more letters!
This seems corroborated by Julian Schilliger (RICHI on Discord) who created his own model that seems to find similar letters.
Wayne has continued to make his own scans of papyrus. Some have even been segmented; see /hari-seldon-uploads/wayne-paths/ on the data server.