Which annotation method is best for phase 1?

[kyrieb-ekat]

Method A:

• three classes: any text, musical symbols (neumes, clefs, custos, divisio), stafflines
• try to get good performance from the start of the process for classification of detected objects to make downstream training in the interactive classifier simpler

Method B:

• two classes: musical block and chant text. All else is ignored.
• extract musical stuff; musical stuff then proceeds to a substep for staffline removal
• limited initial time spent on annotating, greater downstream time spent in the IC

[fujinaga]

My ideas for Method B include train YOLO after staff removal, to recognize neumes (as one class) instead of connected component analysis; before IC. IC would then classify different types of neumes. Alternatively, at least of square (and hufnagel?) notation, the YOLO can recognize individual neumes (punctum, podatus, etc.); and use IC as an editor.