Exciting news from Scroll 5!
For the past several months, our team and community have been working hard on ink detection and segmentation of P.Herc. 172 from the Bodleian Libraries at Oxford (our Scroll 5). We’re excited to share some incredible progress!
The first revelation from P.Herc. 172 was the appearance of visible ink. Unlike most Herculaneum scrolls scanned to date, the ink in this scroll is evidently more dense, appearing sometimes by eye in the scan. This has been a huge boon to our ability to quickly bootstrap ink detection—not unlike our efforts in Scroll 1 during our initial Grand Prize run. Take a look at all this ink!
It’s clear that while there’s plenty of letters, we’ve got some work to do on segmentation. The disconnected columns of text are symptomatic of ‘sheet jumps’, where the segmentation line has accidentally switched to a wrap adjacent to it rather than continuing along its current wrap. As we continue to improve our automatic segmentation pipelines, what we see in this image will instead be large and contiguous rows and columns of beautifully penned Greek.
While work is ongoing, papyrologists at the University of Oxford have been able to gain several preliminary insights into this scroll and its contents.
The scroll offers hints pointing to its likely author being our favorite philosopher in residence: Philodemus. There’s a bit of early evidence that points us in this direction: the letter-forms present in this book suggest it was written sometime in the first century BCE and are of similar shape to handwriting found in other books attributed to him, most of the books found thus far in the library are his, and the word ἀδιάληπτος (‘foolish’) found in this text is characteristic of his writing.
Some other words located so far include διατροπή (‘disgust’), φοβ (‘fear’), and βίου (‘life’). Additionally, the book offers some clues to indicate that it’s likely a finished book, and not an “in-progress” scroll as some that have been found within the library are. For example: this highlighted character in red is not a Greek letter, but a common symbol used at the end of a line to give a justified right-hand margin.
It’s clear from the images that we’ve got a mountain of text left to recover as the technical methods continue to improve. Scroll 5 is particularly promising as its uniquely visible ink offers us a chance to generate more accurate labels with respect to the underlying scroll features.
Perhaps even more promising are the features it shares with the other scrolls in which we have yet to locate ink. Scrolls 1 through 4 each individually contain features that are not common in the others - besides that they are constructed of papyrus, and they all contain the characteristic fibers we’ve come to know. Scroll 5, however, contains a little bit of all of these features. Very rarely are the perfectly flat and clean surfaces of Scroll 1 (P.Herc. Paris 4) present, but it’s full of sections that could easily be mistaken for any of our other scrolls.
It shares this “greasy” or “bubbly” texture that dominates the surface of Scroll 3 (P.Herc. 332). At first glance, other than the obvious intensity difference, it’s hard to differentiate the surfaces at all.
It shares with Scroll 4 (P.Herc. 1667) the plate-like surfaces that are tantalizingly similar to Scroll 1’s “crackle” feature, and have continued to confuse would-be ink hunters for the better part of two years.
Scroll 5’s greatest gift might be its potential ability to operate as a “Rosetta Stone” for ink detection into other scrolls. There is however much work to be done before we get there.
Even though this scroll contains visible ink, it is mostly concentrated towards the outer wraps of the scroll, and ink detection becomes increasingly difficult towards the center.
Scroll 5 segmentations are currently relatively sparse, and its surface features extend as well to its overall volumetric appearance, meaning volumetric fiber or surface labels could prove to be incredibly valuable.
Improvements to either segmentation or ink detections will independently yield significant improvements towards the readability of this scroll. To facilitate this effort, today we are releasing a number of large segmentations in Scroll 5. These were generated through Thaumato and through this past years FASP submission from @Waldkauz and @bruniss, which we’ll call “Surface Traces”. The surface traces are located on the data server here, and the Thaumato surfaces are here.
Although these are not as high quality as some of the meshes we have seen in Scroll 1, they contain many large regions of high- quality surfaces. Currently the meshes from the Surface Tracer are downsampled, we are working on rendering the full resolution ones. The downsampled versions function well using the Grand Prize TimeSFormer model but can struggle with the i3d model from the First Letters Prize.
A few of the surface traces are quite large, this one for example encompasses what appears to be almost half the scroll width!
Our belief is that by utilizing the visible text from this scroll to generate performant ink detection models within Scroll 5, that these shared features will enable us to finally have what we’ve all been working towards this past year: fully generalizable ink detection and representations (and a whole ‘lotta text).
We’re very thankful to our partners at Oxford University and the Bodleian libraries for their work and to our Challenge community who continues to inspire us to do everything we can to read every scroll!
Happy ink hunting!
So you'll be scanning 10 scrolls this year?