Is glyph selection in 5.9 fundamentally broken?

[jandyman]

I have somewhat overcome all the documentation difficulties with training for a new font, though it is fuzzy. However, even after reviewing all the documentation I can find for glyph selection - which is necessary for creating the data for training - I cannot get things to behave. Both lasso selection and Cmd-click seem to behave in psycho ways. Is this known bugs, or is the logic so unobvious and undocumented that I can't figure out how it works? Lasso selection creates both black and red highlighted areas, cmd-click in the unselected (black?) areas to add the missing sections appears to just reset the selection, and in weird ways. What is going on?

[hbitteur]

@jandyman
The various glyphs get available at the end of the SYMBOLS step (even if a few of them are created in earlier steps).
So, make sure you have processed your score until at least the SYMBOLS step, to easily grab your glyphs.

[jandyman]

@hbitteur I don't think you understood my question. I have glyphs that have not been assigned to anything at the end of the analysis process for a score. I want to select them and assign them to an inter, so that they will end up in the repository when I "Sample Book Symbols", and I can subsequently attempt to train Audiveris on this unfamiliar music font. For such purposes, I want to accurately select the entire glyph, otherwise any subsequent training will be counter productive. But when I try to select the glyph, often only part of the glyph is actually selected. Trying cmd-click to add the other areas of the glyph isnt' working. Lassoing often selects only part of a glyph, and then of course I can't fix it with cmd-click. Interestingly. I tried another score which has much poorer scanning quality, such that the black/white determination is more noisy, and this score did not seem to have a similar selection problem. The "cleaner" score was actually worse with regards to only selecting parts of a glyph!

If you like, I can send the score in question which I just cannot get to work and describe the glyphs I cannot see to select properly, then perhaps it will make more sense. I hope I have described the problem this time in a way that is easier to understand.

[hbitteur]

Yes, post an example if you can

[jandyman]

@hbitteur Here is an example zipped up. There is the original pdf, two png examples of the "red/black" selection issue, and the omr file in case it is helpful. I should note that this was the file I worked on when I encountered all the training issues I posted previously about, but had in fact subsequently succeeded in training it to detect the "flat" symbols more accurately. If you use a default training set, it probably won't detect as many of these symbols. But I don't think it has anything to do with this issue, I experienced the same glyph selection issues both before and after my semi-successful training efforts.
example.zip

[hbitteur]

@jandyman
OK, I understand now.
A lasso will select several things embraced by the lasso rectangle.

In the precise cases (ex1.png and ex2.png) what is displayed is:

• The "main" glyph in red
• Temporary glyphs, such as stem seeds, may still be there and are displayed in black, perhaps on top of the "main" glyph.

The "main" glyph is what is relevant for your assignment.

Now, you can save your book and reload it. The "temporary" black glyphs will be gone and you won't be disturbed in your manual selections and assignments.

I agree this can be confusing.
I should discard the temporary glyphs as soon as they are no longer useful. I'll put this on the TODO list.

[jandyman]

@hbitteur
OK, that's progress. My summary of what you said would be that the black areas are "on top of" the red areas, if you were to hide them, the red area would display the whole glyph properly. And things do seem to behave that way. Is this correct?

But I'm still having difficulty getting training to work, in fact, I seem to have taken a step backwards. Let me tell you what I've done:

1. I select 5 quarter rest glyphs and assign them to quarter rest inters.
2. I sample the book symbols
3. I browse the book repository and see my assignments, they look good.
4. I push to the global repository.
5. I close the book repository dialog, saving when it asks me.
6. I open the global repository and see my assignments there
7. I select Train Classifier. The dialog opens up.
8. I press Select, then Train. It seems to train, it takes a while.
9. I press the two Test buttons, everything seems perfect.
10. I go back to the original book to see if the training has helped anything. I reset the book to binary
11. I re run the analysis. But Audiveris still doesn't select any of the quarter rests properly.

Is this the right process, or did I get a step wrong? Should Audiveris be picking up the quarter rests more successfully? Would be helpful for me to send any files to see what might be going on?

This is so close to working right, and I get a number of scores with the music font in question, so if I could solve this, I'd be in really good shape. I've been told that this font is a Finale font, just not a common one. And BTW, I've tested a couple of commercial options for score detection (not the multi hundred dollar ones), and they haven't done any better. So I'm very motivated to get Audiveris to work with this font.

[hbitteur]

@jandyman
Your process is the right one, everything should work.

I'd suggest that you simply re-open the global repository.
There you should find the 5 quarters you have pushed.
Click on every one of them, this triggers the glyph classifier which should tell you how it classifies the sample at hand and with which grade.
It should correctly recognize your samples.

However: there are plenty of other quarter samples in the repository, perhaps your 5 ones are too few to really modify the model. And indeed "your" quarter rests don't look like standard ones...

I'm not sure.
So, please do this, and share the result. Then we'll decide what to try next.

[hbitteur]

OK, that's progress. My summary of what you said would be that the black areas are "on top of" the red areas, if you were to hide them, the red area would display the whole glyph properly. And things do seem to behave that way. Is this correct?

This is correct!

[jandyman]

@hbitteur
I'm not sure what I'm looking at when I trigger the glyph classifier, it seems to display information that seems non-sensical to me. Here is a typical example:

[hbitteur]

Well, my understanding is that (upper right) you have selected the 3rd sample in your row of 9 samples.
And in the Glyph Classifier board, we can read that the "best" shape assigned is a "15" quindicesima with a very low confidence: 0.0014

Which is way too low to assign a shape to a glyph. IIRC, it needs at least 0.10 to simply consider it, and finally at least 0.5 (contextual grade) to keep it.
We are far from that!
And consequently, these quarter rests won't be picked up when you re-process your images.

I'd suggest two things: augment the number of your quarter samples, train the model longer.
Typically, when I train the model from scratch, I use 2 training sessions one after the other.
This results in a total of about 1000 epochs.

[hbitteur]

Important question: Did you train from scratch (by first resetting the model) or did you incrementally run it?

[jandyman]

@hbitteur
I don't know the difference between training from scratch or incrementally, or how to reset a model. I've tried my best to understand how all this training stuff works from the documentation, but the docs are relatively sketchy. (I understand how difficult it is to keep docs up to date - I'm a SW developer too).

I can create more quarter samples and train the model longer. The AI searches I did tended to indicate that 5 samples of a glyph wouldn't be too bad.