PROPOSED "RULES" FOR MULTIPLE-CLASSIFIER OMR
Don Byrd, rev. 9 Nov. 2005

Implementing a multiple-classifier system requires (a) procedural rules, or at
least rules that are precise enough that they can be made procedural, and (b)
registration (a.k.a. alignment) of output from the various classifiers so
direct comparisons can be made. Here are some proposed rules for such a
system. These are much too vague, besides being based on very little data;
they're intended, first, to give an idea of the kind of rules needed (aside
from precision), and second, as a starting point for further experimentation
with the systems.


SE = SharpEye 2.63; PS = PhotoScore Professional 3 for Windows; SS =
SmartScore Pro 3 for Windows. Based on experience in the SFStudy to date,
plus comments from others (identified as such):

1. In general, SE is the most accurate.

2. PS often misses C clefs, at least in one-staff systems (e.g., OMR Quick-Test).

3. For text (excluding lyrics and standard dynamic-level abbreviations like
	"pp", "mf", etc.), PS is the most accurate (e.g., TP 12).

4. Resolution (300 vs. 600 dpi) affects PS far more than the others--though
	sometimes it does much better at high resolution, sometimes much worse.

5. PS at low resolution makes the most mistakes on notes; it also adds the
	most extraneous augmentation dots (e.g., Quick-Test at 300 DPI).

6. SS is by far the best at hairpins; SE is worst (e.g., QS, Quick-Test,
	Beethoven).

8. SS usually/always finds pedal downs; PS and SE never find them.

9. SS is the best at recognizing non-slur articulation marks; PS is the worst
	(e.g., TP 12).

10. SS is the only one that ever recognizes 1st and 2nd endings (e.g., TP 12).

11. SE is the worst at recognizing beams, i.e., the most likely to miss them
	completely (e.g., TP 12 & 13).
	
12. PS is the worst (though it's still not bad) at recognizing diatonic pitch
	(line/space position), especially at low resolution.

13. PS is the only one that ever recognizes octave signs (e.g., QS, Quick-Test).

14. SS is the only one that ever recognizes arpeggio signs (e.g., QS).

15. SE is the only one that ever recognizes trill markings (e.g., Mozart).

16. SS is the only one that ever recognizes cross-staff beams (e.g., QS).

17. SE is the only one that ever recognizes grace notes (e.g., TP 17 & 21).

-----------------------------------------------------------------------------

The first version of these rules (2 June) was:

1. In general, SE is the most accurate.

2. PS never recognizes C clefs in one-staff systems.

3. For text, PS is the most accurate; SS is next; SE is the least accurate.

The idea that SE is least accurate with text was based on tests in which
we inadvertantly had its text recognition turned off :-) .