We need Mothra to be excellent at classifying text-stafflines-music and providing bounding boxes

[kyrieb-ekat]

[kyrieb-ekat]

Currently in the version 1 model for the "big annotation" phase things are promising; the initial set of training data skewed toward quadratic neumes and gothic textualis, but the model did still perform well on unseen types of document layout(s) and notations, as well as on pages without stafflines at all. Some items do worse than others, such as some types of St Gall documents with visible ruling lines, which are guessed as stafflines, and Hispanic pages (Leon Ms 8) where both text and music are not often detected even though the items that are classified are usually correct.

Ensuing items that are being newly annotated and corrected are targeted at addressing the staffline gap (we have significantly fewer staffline samples for our ground truth than anything else; for exact numbers, please see the discussion on this in the mothra repo) and the adiastematic gap; additional pages in both St Gall, Aquitanian, and Hispanic have been added, and previous Beneventan and Messine samples are being corrected.

[annamorphism]

the title of this issue doesn't seem to reflect an issue to be resolved. can we rephrase what the problem is here?

[kyrieb-ekat]

Good shout, fixed! I'll be zipping about making a bunch of sub issues shortly as well