Audiveris option tweaks for orchestral scores

[Alterrien]

Hey there,
I'm trying to get Audiveris to work on a full orchestral score. The PDF in itself seems to be a direct export from an engraving software, so it should be fairly clean to work with. However, it seems like Audiveris struggles a lot with the smallness of all parts of the score.

You can get the PDF directly from IMSLP there: https://vmirror.imslp.org/files/imglnks/usimg/d/de/IMSLP741670-PMLP690923-THE_CLOUD_MESSENGER_FINAL.pdf.

One thing I had to tune is to set the min interline value to be much lower. It correctly detected the interline of 8, but could not apply it as it was lower than that setting. The error shown in the "should we set this page as invalid" was a bit misleading. See the logs below.

[IMSLP741670#7] IMSLP741670#7 With an interline value of 8 pixels, either this sheet contains no staves, or the picture resolution is too low (try 300 DPI).
[IMSLP741670#7] Beam  guessed height:  4 -- 0.50 of 8 interline
[IMSLP741670#7] Beam measured height:  7 -- 0.83 of [2..8] range at 59% of needed quorum
[IMSLP741670#7] No reliable beam height found, guessed value: 4
[IMSLP741670#7] Scale{ interline(8,8,9) line(1,1,2) beam(4 extra)}
[IMSLP741670#7] End of sheet step SCALE

With this issue fixed, I can get a bit further in the pipeline, but it quickly fails at the BEAMS step with a nasty java.lang.RuntimeException: org.audiveris.omr.math.Line$UndefinedLineException: Not enough defining points : 1

Would you have an idea on both how to fix my particular issue, and if there is a set of options I can set for Audiveris to better handle my score?
Thanks in advance!

[hbitteur]

@Alterrien
The problem is the very low resolution of the input image.
I tried the first sheet which contains some music staves, that's sheet number 7.

The scale step computed a interline value of 8 which is lower than the accepted minimum (11).
So I bypassed the protection by manually setting the constant minInterline in the class ScaleBuilder, from 11 down to 8.
The SCALE step was now OK, even giving me an extrapolated value for beam thickness of 4 (which is correct).

The problem is that the engine now operates in some abnormal land.
The next error occurred in the BEAMS step, when the engine tried to consider the big "M" in the text "The Cloud Messenger" as a candidate spot for a beam!! The reason is the music scale is so small that letters are comparatively very big.
Here again, after some debugging, I discovered that the constant coreSectionWidth which defines the minimum horizontal length for a core section on a beam top or bottom border, got down to the integer value 1 (it is defined as 0.15 times the interline 8)., which means a length of 1 pixel. And this is not enough to define a horizontal line, which requires at least 2 points. Hence the error message.

Here again I somehow cheated, by setting this constant from 0.15 to 0.5, and the BEAMS step no longer crashed.

I was able to get the whole sheet transcribed, but many mistakes remain that would need manual correction by the end user.
So, I don't know if it's worth to fight with this input.

For example, here is a sharp ?

And here, see the head, the augmentation dot, and a slur which appears rather thick (still by comparison).

Is there any way to get this score with a higher resolution?

[hbitteur]

Update:
I just pushed a bug fix for very low resolution.
There is no more need to manually change coreSectionWidthin class BeamsBuilder to survive the BEAMS step.

[Alterrien]

Hi Hervé, thanks a lot for all your feedback!
I don't think I'll be able to get a higher resolution file, specifically since this is a score that I'd like to extract split parts. However, since this is basically an export from an engraving software, I'm guessing the PDF should have vector graphics instead of a picture, so I'm guessing we should be able to artificially increase the resolution of the PDF? I'll try to look into that.

[hbitteur]

Very good idea, assuming of course that the PDF uses vector graphics.
Audiveris now reads PDF files via the Apache PDFBox software, perhaps you could have a look there.

[hbitteur]

Just my two cents:
You might be interested by this handbook section enlarging low resolution scans contributed a couple of years ago by Ram Kromberg

[Alterrien]

Thanks for the tip ! I did a very simple test using imagemagick to upscale the pdf to 600dpi (and convert it to png). I can send you the image next time I'm on my computer. Audiveris fared much better, going through all the steps successfully and correctly figuring out a bunch of the first page, but it still feels like I was running against the limit of the system. Most notably the resulting PNG went above the pixel size limit of 20M that I had to go over, and there was still a fair amount of misidentifying glyphs and some poor dealing with quirks of orchestral notations. Le mar. 10 déc. 2024, 18:26, Hervé Bitteur ***@***.***> a écrit :

…

Just my two cents: You might be interested by this handbook section enlarging low resolution scans <https://audiveris.github.io/audiveris/_pages/advanced/improve_input/#enlarging-low-resolution-scans> contributed a couple of years ago by Ram Kromberg — Reply to this email directly, view it on GitHub <#762 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACRNXCMQDDRJE6TIYGLNBDD2E4P3NAVCNFSM6AAAAABTBEJPF2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKMZSGM2DSNBQGY> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[Alterrien]

Hey again,
Here's the PNG, converted from PDF through convert -density 600 -depth 8 -quality 85 cloud-messenger-7.pdf cloud-messenger.png. I haven't been able to go on my GPU-powered PC so I haven't tried the waifu2x option to compare, I'll report back.
I'm just getting my bearings on using Audiveris so I don't know if the amount of mistakes to fix is on par for the course, but hopefully with some more experience I'll be able to go through those pages quicker.

[hbitteur]

Quality is much better this time.
The amount of manual fixes is on par with other scores, even though we have more musical symbols in such an orchestra score.

FYI, I'm attaching TheCloudMessenger.zip wich contains the resulting .omr (which I cannot upload directly on this forum)
TheCloudMessenger.zip