Genesis

As opposed to OCR (Optical Character Recognition) for which reference material has existed for a long time and has boosted the development of industrial projects, there is yet no similar data publicly defined for OMR.

Here below are recent projects or events that somehow triggered this new OMR dataset project.

Audiveris OMR

Audiveris, just like any other OMR software, needs score and symbol samples in order to train its internal symbol classifier.

Until version 4, Audiveris was only able to process a score image in memory and deliver its symbolic information using MusicXML format. The training data consisted only in a few glyph images, carefully selected by hand.

Starting with version 5, Audiveris now stores its internal data into ".omr" project files, using a public data model. Any program can thus read this data and retrieve precise information about any symbol (its name, the pixels it is made of, its precise location in score image, etc). Manual checking is still needed to review and possibly discard some symbols but this processing gets more and more efficient, since the classifier is continuously trained on more and more valid data and thus delivers information which requires less and less validation effort. We thus can speak of a form of "boot-strapping".

ZHAW Deep Score project

[TODO: More information is needed on this Deep Score project]

Since mid 2016, ZHAW university ( Zürcher Hochschule für Angewandte Wissenschaften, in English: Zurich University of Applied Sciences) collaborates with Audiveris for its Deep Score project.

Audiveris provides a working OMR framework so that ZHAW can test in situ the use of deep learning techniques applied to OMR.

In March 2017, symbol context images (score sub-images centered on symbol center) were used to train a classifier with very good results reported at symbol level.

Salzburg Classical Music Hackday

On April 8-9 2017 in Salzburg, within the Classical Music HackDay, a hack was proposed by Audiveris and MuseScore to set up an OMR dataset from synthetic score images generated from MuseScore symbolic data.

MuseScore has thousands and thousands of scores in symbolic format, from which score images can easily be drawn with precise knowledge about every symbol shape and location. We were well aware that these "artificial" images would be "too perfect" as compared with the kind of ordinary scans that are submitted to OMR. However, we felt that we had the opportunity to hack the production of some OMR data that might be usefully tested for OMR classifier training. So we did it.
See the presentation made about Audiveris and hack proposal with MuseScore.

Time was awfully limited, but Sanu Pulimootil (from Bavarian State Library) volunteered to take over the production of deteriorated images, and finally we could go through all the needed steps to exercise the framework on a toy set of symbols. We did produce some score images annotated with symbol information, try various means of image deterioration, extract the corresponding symbol sub-images, then define and train a convolutional neural network on this data.
See the presentation made about early hack results.