OMR Training
The github page mentions how a specific branch of musescore was used to generate the training data files for the OMR learning project. When I build that musescore branch & try to use it:
mscore file.mscz -o file.png -E
The resopnse is that the "-E" flag is unknown.
How do I get the XML files that have annotations in them to match the image files (as per the OMR-dataset project)?
Comments
-e (lowercase) for "experimental"?
Hi, you need to compile this branch https://github.com/musescore/MuseScore/tree/imeta
And it's indeed -E or --emeta
Hello, I'm Hervé Bitteur, owner of Audiveris OMR project. Perhaps you already hit https://github.com/Audiveris/omr-dataset-tools which is an attempt to create a large OMR training dataset, something we launched with MuseScore in Salzburg a few months ago. But MuseScore guys have plenty of other tasks to carry...
Would you be interested in contributing to this project? The C++ code on MuseScore side is nearly usable, I have formalized a dozen of needed modifications as described on https://github.com/Audiveris/omr-dataset-tools/issues
On Audiveris Java side, I'm ready to design several OMR classifiers and pre-train them on MuseScore output.
Just tell me.
/Hervé
Easier said than done, it seems. I'm now trawling my way through the difficulties presented by that branch. At least your message reminded me that I hadn't changed to the right branch.
In reply to Easier said than done, it… by Jax Jondarr Gibb
Hi again. I posted a reply 2 days ago, reportedly "waiting for some moderator", but it is still not showing up. Perhaps, things will go better this time.
I'm Hervé Bitteur, owner of Audiveris OMR project https://github.com/Audiveris. It includes a sub-project named OMR-dataset, launched with MuseScore during Salzburg hackday a few months ago.
The code to populate this OMR dataset from MuseScore still needs some minor bug fixing, see the issues flagged with "musescore" on https://github.com/Audiveris/omr-dataset-tools/issues
Would you be interested in completing this task (it requires C++ competency)?
On my side (Audiveris Java), I'm ready to take care of the design of new classifiers and their training on musescore output.
Please tell me
/Hervé
In reply to Hi again. I posted a reply 2… by Hervé Bitteur
Hervé,
fantastic to hear from you, & it's also good to hear that there is an owner in this area. Although I'm not a C++ dev, I'm slowly getting through the build. The chief problem is that the result may not be reproducible, because I'm working in windows without an IDE. Not ideal.
As for the training, here is my situation: as bagpipe players, we share a lot of music around where someone has a print-out of a tune - often produced by software (that is, neat & regular) - from someone else, but it takes effort to transcribe it & then make local modifications. I had hoped to use Musescore trained specifically on the limited range of bagpipe music to at least be able to produce an editable form. Looking at the training data, I thought this would be relatively simple - a matter of generating the notes & embellishments in the right format. Failing that, I thought I would be able to take samples from my large collection (in BWW), convert these to MusicXML (or just import them into MuseScore), then generate the training data as full scores.
There are a lot more moving parts in this process than I had envisaged when I decided to take this on.
I guess, fundamentally, I'll stay in touch, struggling through the MuseScore build (I'm not the only one, I understand), & some day one of us will be able to PR back into the repo.
Jax
In reply to Hervé,… by Jax Jondarr Gibb
If needed you might find some direct help for the build process on IRC #musescore on freenode.net and of course here https://musescore.org/en/developers-handbook/compilation