Ground Truth Status Tracking

[kyrieb-ekat]

============================================================

DATASET GROUND TRUTH STATISTICS

============================================================

as of 19 March 2026

Images: 17
YOLO .txt files: 17
Total annotations: 6723

Per-class breakdown:
Text (0): 1733 ( 25.8%)
Music (1): 4177 ( 62.1%)
Staves (2): 813 ( 12.1%)

============================================================

Hufnagel

Antiphonal_12v_hfngl.jpg
Antiphonal_1v_hfngl.jpg
Antiphonal_44v_hfngl.jpg

NZ-Wt MSR-03 013r.png
NZ-Wt MSR-03 065r.png
NZ-Wt MSR-03 109v.png

Square/quadratic

CH-Fco2_009v_sqr.jpg.jpg
CH-Fco Ms. 2_006r copy.jpg
CH-Fco Ms.2 108v.jpg
CH-Fco2_002v_sqr.jpg

F-Pn Latin 15181 169v bigger.png

Gothic

D1161_210546_gthc.jpg
D1161_210404_gthc.jpg
D-KNd 1161 032r.jpg

Messine

CH-P 18 p.100.jpg
CH-P18_061_mssn.jpg.jpg
CH-P18_288_mssn.jpg

Square	Hufnagel	Messine	Gothic	Hispanic
CH-Fco	Antiphonal / NZ-Wt	CH-P 18	D1161 / D-KNd	---
F-Pn Latin 15181	---	---	---	---

Performance w/ my small terrible proofofconcept script+model

Performance: Text mAP@50 = 0.77, Music mAP@50 = 0.67, Staves mAP@50 = 0.02
To see predictions on unrelated manuscripts and unseen neume script styles (adiastematic!), look here.

Coming Up

• We need a lot more staves
• We can keep generalising, but we need to annotate across types in sync

[kyrieb-ekat]

Alright I got morbidly curious and screenshotted a page from the internet archive of a copy of the LU and tested v1 mothra (models/v1/yolov8m_17images_7568boxes.pt; never seen print before) on a print page and if you don't count stafflines it did great. Counting stafflines... uhhh.... we still have some work to do clearly. Putting this here as a project tracking/note. @CThierrin any specific samples you can pull from for stafflines to add for re/training will be a huge help.