Mothra annotation styleguide (ever evolving)

[kyrieb-ekat]

Setting this up as a running place for observations on how we're annotating our mothra ground truth. @JoyfulGen this will be pretty much just a way for us (and whoever helps us out with annotation) to keep track of what we're doing per phase and per new model/tweak.

METHOD A ANNOTATION GUIDELINES

Current phase: 1

Objective: general detection

Guidelines a/o May 2026, modelv1.1 (from YOLOv8, proof of concept findings)

1. Initials should be in their own annotation box, not included with the ensuing letters/rest of the word. So: S is in one box and alve is in another

2. miniscules or sufficiently nearby majiscules should be included as text (in below: all one annotation box)

3. for now we're ignoring decoration including absolutely gigantic initials, lombards, and flourishes. If it's an in-text flourish, include it.

YES:

NO:

4. If a staffline kinda peters out in the middle of the page, but continues a little further along, box the whole thing as ONE SINGLE bounding box; do not break up!

5. some gothic and messine pages will mark an f-clef as a dot on the red line. Make sure to annotate those as "music" for now. Things on a staffline that would count as music = music.

6. if there is a large gap between syllables in a word, don't group them as one word, group them by space. So, if your thumb fits between the two words: two boxes. If it doesn't: one big box.

7. include divisio (no matter how wonky/hairline) as an annotated "music" element.
8. include abbreviation symbols in a text box (ex: ā).
9. do annotate marginalia where present

[kyrieb-ekat]

METHOD B ANNOTATION GUIDELINES

Very similar to the above, but you will instead only be annotating for two classes! Chant-text and music blocks.

Below is a square notation manuscript with chant text grouped as a line, and the music block including stafflines and musical glyphs (clefs, neumes, custodes). All other items will be ignored. Majuscule letters may be included; initials shall not. Rubrics will be ignored. @fujinaga , is this correct?

[kyrieb-ekat]

REGION ANNOTATION GUIDELINES

We’re going to use mothra-annotator for this, and label the “regions we want” as text and “regions we don’t want” as staves.

Regions we want: text+music block (all in one box, a la method B), the whole writing space (one big box around the central writing rectangle/where the writing is happening, excluding genre cues and initials; furthest top, left, bottom, and right of where there’s black ink), marginalia (music).

• it’s possible that a line region will overflow beyond the edges of the writing space boundary--this is ok.
• overlapping ok

Regions we don’t want: initials, marginalia (text), decoration.