Optical Music Recognition Datasets

This repository contains a collection of many datasets used for various Optical Music Recognition tasks, including staff-line detection and removal, training of Convolutional Neuronal Networks (CNNs) or validating existing systems by comparing your system with a known ground-truth.

Note that most datasets have been developed by researchers and using their dataset requires accepting a certain license and/or citing their respective publications, as indicated for each dataset. Most datasets link to the official website, where you can download the dataset.

If you are interested in Optical Music Recognition research, you can find a curated bibliography at https://omr-research.github.io/.

Overview

The following datasets are referenced from this repository:

If you find mistakes or know of any relevant datasets, that are missing in this list, please open an issue or directly file a pull request.

Tools for working with the datasets

A collection of tools that simplify the downloading and handling of datasets used for Optical Music Recognition (OMR). These tools are available as Python package omrdatasettools on PyPi.

Handwritten Online Musical Symbols (HOMUS)

Official website: http://grfia.dlsi.ua.es/homus/

Summary: The Handwritten Online Musical Symbols (HOMUS) dataset is a reference corpus with around 15000 samples for research on the recognition of online handwritten music notation. For each sample, the individual strokes that the musicians wrote on a Samsung Tablet using a stylus were recorded and can be used in online and offline scenarios.

Scientific Publication: J. Calvo-Zaragoza and J. Oncina, "Recognition of Pen-Based Music Notation: The HOMUS Dataset," 2014 22nd International Conference on Pattern Recognition, Stockholm, 2014, pp. 3038-3043. DOI: 10.1109/ICPR.2014.524

Example:

Remarks: The original dataset contains around 20 artifacts and misclassifications that were reported to the authors and corrected by Alexander Pacha.

Universal Music Symbol Collection

Official website: https://github.com/apacha/MusicSymbolClassifier, Slides

Summary: A collection of various other datasets which combines 7 datasets into a large unified dataset of 90000 tiny music symbol images from 79 classes that can be used to train a universal music symbol classifier. 74000 symbols are handwritten and 16000 are printed symbols.

Scientific Publication: Alexander Pacha, Horst Eidenberger. Towards a Universal Music Symbol Classifier. Proceedings of the 12th IAPR International Workshop on Graphics Recognition, Kyoto, Japan, November 2017. DOI: 10.1109/ICDAR.2017.265

Example:

CVC-MUSCIMA

Official website: http://www.cvc.uab.es/cvcmuscima/index_database.html

Summary: The CVC-MUSCIMA database contains handwritten music score images, which has been specially designed for writer identification and staff removal tasks. The database contains 1,000 music sheets written by 50 different musicians. All of them are adult musicians, in order to ensure that they have their own characteristic handwriting style. Each writer has transcribed the same 20 music pages, using the same pen and the same kind of music paper (with printed staff lines). The set of the 20 selected music sheets contains music scores for solo instruments and music scores for choir and orchestra.

Scientific Publication: Alicia Fornés, Anjan Dutta, Albert Gordo, Josep Lladós. CVC-MUSCIMA: A Ground-truth of Handwritten Music Score Images for Writer Identification and Staff Removal. International Journal on Document Analysis and Recognition, Volume 15, Issue 3, pp 243-251, 2012. DOI: 10.1007/s10032-011-0168-2

Example:

MUSCIMA++

Official website: https://ufal.mff.cuni.cz/muscima

Current development: https://github.com/OMR-Research/muscima-pp

Summary: MUSCIMA++ is a dataset of handwritten music notation for musical symbol detection that is based on the MUSCIMA dataset. It contains 91255 symbols, consisting of both notation primitives and higher-level notation objects, such as key signatures or time signatures. There are 23352 notes in the dataset, of which 21356 have a full notehead, 1648 have an empty notehead, and 348 are grace notes. Composite objects, such as notes, are captured through explicitly annotated relationships of the notation primitives (noteheads, stems, beams...). This way, the annotation provides an explicit bridge between the low-level and high-level symbols described in Optical Music Recognition literature.

Scientific Publication: Jan Hajič jr., Pavel Pecina. The MUSCIMA++ Dataset for Handwritten Optical Music Recognition. 14th International Conference on Document Analysis and Recognition, ICDAR 2017. Kyoto, Japan, November 13-15, pp. 39-46, 2017. DOI: 10.1109/ICDAR.2017.16

Example:

Remarks: Since this dataset is derived from the CVC-MUSCIMA dataset, using it requires to reference the CVC-MUSCIMA as well.

MUSCIMA++ Measure Annotations

Website: https://omr-datasets.readthedocs.io.

Summary: Based on the MUSCIMA++ dataset, a subset of the annotations was constructed, that contains only annotations for measure and stave recognition. The dataset has some errors fixed that version MUSCIMA++ 1.0 exhibits and comes in a plain JSON format, as well as in the COCO format.

This dataset was created by Alexander Pacha and can be directly downloaded from here.

Example:

DeepScores

Official website: https://tuggeluk.github.io/deepscores/

Summary: Synthetic dataset of 300000 annotated images of written music for object classification, semantic segmentation and object detection. Based on a large set of MusicXML documents that were obtained from MuseScore, a sophisticated pipeline is used to convert the source into LilyPond files, for which LilyPond is used to engrave and annotate the images. Images are rendered in five different fonts to create a variation of the visual appearance.

Scientific Publication: Lukas Tuggener, Isamil Elezi, Jürgen Schmidhuber, Marcello Pelillo, Thilo Stadelmann. DeepScores - A Dataset for Segmentation, Detection and Classification of Tiny Objects. ICPR 2018. 2018. https://arxiv.org/abs/1804.00525

Example:

DoReMi

Official website: https://github.com/steinbergmedia/DoReMi/

Summary DoReMi includes around 6432 images of sheet music with nearly a million annotated objects. Each object on the page is annotated with category labels from 94 different classes. Given that this dataset is generated using Dorico (a music notation software from Steinberg), we had the chance to retrieve more musical information, but also different types of data. Using Dorico we generate 5 types of data that could possibly be used in different steps of OMR. These data are images of scores (binary or colour), the musicXML file, XML files with metadata such as bounding boxes of each object and pixel mask information together with musical information on how noteheads, beams stems and slurs are linked but also their duration beats, onset beats, pitch octave, staff id. Most of the images include one system per page; depending on the number of voices, they will have one or more staves per page.

Scientific Publication: Shatri, Elona, and György Fazekas. DoReMi: First glance at a universal OMR dataset. WoRMS 2021, 2021. https://arxiv.org/abs/2107.07786

Example

<Node>
<Id>504</Id>
<ClassName>noteheadBlack</ClassName>
<Top>466</Top>
<Left>2236</Left>
<Width>28</Width>
<Height>24</Height>
<Mask>0:12 1:11 0:14 1:16 0:10 1:19 0:8 1:21 0:6 1:22 0:5 1:24 0:3 1:25 0:2 1:26 0:1 1:27 0:1 1:27 0:1 1:54 0:1 1:27 0:1 1:26 0:2 1:26 0:2 1:25 0:3 1:24 0:4 1:23 0:5 1:22 0:6 1:21 0:8 1:19 0:10 1:16 0:13 1:13 0:18 1:6 0:16 </Mask>
<Outlinks>541 559 571</Outlinks>
<Data>
<DataItem key="duration_beats" type="float">0.333333</DataItem>
<DataItem key="onset_beats" type="float">47.666668</DataItem>
<DataItem key="pitch_octave" type="int">4</DataItem>
<DataItem key="midi_pitch_code" type="int">69</DataItem>
<DataItem key="normalized_pitch_step" type="str">A</DataItem>
<DataItem key="dorico_event_id" type="int">1120</DataItem>
<DataItem key="staff_id" type="int">0</DataItem>
</Data>
</Node>

PrIMuS

Official website: https://grfia.dlsi.ua.es/primus/

Summary: The Printed Images of Music Staves (PrIMuS) contains the 87678 real-music incipits (an incipit is a sequence of notes, typically the first ones, used for identifying a melody or musical work) in five different formats: As rendered PNG image, as MIDI-file, as MEI-file and as two custom encodings (semantic encoding and agnostic encoding). The incipits are originally taken from the RISM dataset.

PrIMuS has been extended into the Camera-PrIMuS dataset that contains the same scores, but the images have been distorted to simulate imperfections introduced by taking pictures of sheet music in a real scenario.

Scientific Publications:

• Jorge Calvo-Zaragoza and David Rizo. End-to-End Neural Optical Music Recognition of Monophonic Scores. Applied Sciences, 2018, 8, 606. http://www.mdpi.com/2076-3417/8/4/606 (for PrIMuS)
• Jorge Calvo-Zaragoza and David Rizo. Camera-PrIMuS: Neural end-to-end Optical Music Recognition on realistic monophonic scores. In Proceedings of the 19th International Society for Music Information Retrieval Conference, Paris, 2018. http://ismir2018.ircam.fr/doc/pdfs/33.pdf (for Camera-PrIMuS)

Example:

Baró Single Stave Dataset

Official website: http://www.cvc.uab.es/people/abaro/datasets.html

Summary: The Single Stave dataset by Arnau Baró is a derived dataset from the CVC-MUSCIMA dataset and contains 95 single stave music scores with ground truth labels on the symbol level.

Scientific Publication: Arnau Baró, Pau Riba, Jorge Calvo-Zaragoza, and Alicia Fornés. From Optical Music Recognition to Handwritten Music Recognition: a Baseline. Patter Recognition Letters, 2019 (in press). DOI: 10.1016/j.patrec.2019.02.029

Example:

Multimodal Sheet Music Dataset

Official website: https://github.com/CPJKU/msmd

Summary: MSMD is a synthetic dataset of 497 pieces of (classical) music that contains both audio and score representations of the pieces aligned at a fine-grained level (344,742 pairs of noteheads aligned to their audio/MIDI counterpart). It can be used for training and evaluating multimodal models that enable crossing from one modality to the other, such as retrieving sheet music using recordings or following a performance in the score image.

Scientific Publications:

• Matthias Dorfer, Jan Hajič jr., Andreas Arzt, Harald Frostel, Gerhard Widmer. Learning Audio-Sheet Music Correspondences for Cross-Modal Retrieval and Piece Identification. Transactions of the International Society for Music Information Retrieval, issue 1, 2018.

Example:

Sheet Midi Retrieval Dataset

Official website: https://github.com/tjtsai/SheetMidiRetrieval

Summary: This dataset contains the scores for 200 music pieces along with their MIDI representation and query images with they ground-truth alignment.

Scientific Publications:

• Timothy Tsai, Daniel Yang, Mengyi Shan, Thitaree Tanprasert, TTeerapat Jenrungrot. Using Cell Phone Pictures of Sheet Music To Retrieve MIDI Passages. IEEE Transactions on Multimedia. 2020

AudioLabs v1

Official website: https://www.audiolabs-erlangen.de/resources/MIR/2019-ISMIR-LBD-Measures

Summary: The data set provides measure annotations for several hundred pages of sheet music, including the complete cycle Der Ring des Nibelungen by Richard Wagner, selected piano sonatas by Ludwig von Beethoven, the complete cycle Winterreise by Franz Schubert, as well as selected pieces from the Carus publishing house.

Scientific Publication: Frank Zalkow, Angel Villar Corrales, TJ Tsai, Vlora Arifi-Müller, and Meinard Müller: "Tools for Semi-Automatic Bounding Box Annotation of Musical Measures in Sheet Music". Late Breaking/Demo at the 20th International Society for Music Information Retrieval, Delft, The Netherlands, 2019. Download the PDF

Example:

AudioLabs v2

Official website: Download the dataset

Summary: AudioLabs v2 is an extension of the AudioLabs v1 dataset with 24,186 bounding boxes for system measures, 11,143 bounding boxes for stave annotations and 50,651 bounding boxes for staff measures, which where generated with the help of a neural network and the original dataset. Annotations are available in the original CSV format, plain JSON format and COCO format.

Example:

MuseScore

Official website: https://musescore.com/sheetmusic

Summary: MuseScore is a free music notation software and also allows their users to upload their sheet music to their website and share it with others. Currently (Jan. 2018) the website hosts over 340000 music sheets, that can be downloaded as MuseScore file (mscz), PDF, MusicXML, MIDI and MP3.

Publication: https://musescore.org

Example:

MuseScore Monophonic MusicXML Dataset

Official website: https://github.com/eelcovdw/mono-musicxml-dataset

Summary: This dataset contains the IDs to 17000 monophonic scores, that can be downloaded from musescore.com. A sample script is given that downloads one score, given you've obtained a developer key from the MuseScore developers.

Scientific Publication: Eelco van der Wel, Karen Ullrich. Optical Music Recognition with Convolutional Sequence-to-Sequence Models. CoRR, arXiv:1707.04877, 2017. https://arxiv.org/abs/1707.04877

Examples:

Capitan collection

Official website: http://grfia.dlsi.ua.es/

(Freely available for research purposes)

Summary: A corpus collected by an electronic pen while tracing isolated music symbols from Early manuscripts. The dataset contains information of both the sequence followed by the pen and the patch of the source under the tracing itself. In total it contains 10230 samples unevenly spread over 30 classes. Each symbol is described as stroke (capitan stroke) and including the piece of score below it (capitan score).

Scientific Publication: Jorge Calvo-Zaragoza, David Rizo and Jose M. Iñesta. Two (note) heads are better than one: pen-based multimodal interaction with music scores. Proceedings of the 17th International Society of Music Information Retrieval conference, 2016. Download the PDF

Example:

Remarks: This dataset exists in two flavours:

• As raw dataset, which contains only the textual descriptions of the strokes and the images, called Bimodal music symbols from Early notation. This format is similar to the HOMUS dataset.
• As rendered images inside of the Isolated handwritten music symbols dataset. Also refered to as Capitan collection.

SEILS Dataset

Official website: https://github.com/SEILSdataset/SEILSdataset

Summary: The SEILS dataset is a corpus of scores in lilypond, music XML, MIDI, Finale, **kern, MEI, **mens, agnostic, semantic and pdf formats, in white mensural and modern notation. The transcribed scores have been taken from the 16th century anthology of Italian madrigals Il Lauro Secco, published for the first time in 1582 by Vittorio Baldini in Ferrara (Italy). The corpus contains scores of 30 different madrigals for five unaccompanied voices composed by a variety of composers.

Scientific Publication: Emilia Parada-Cabaleiro, Anton Batliner, Alice Baird, Björn W. Schuller. The SEILS dataset: Symbolically Encoded Scores in ModernAncient Notation for Computational Musicology. Proceedings of the 18th International Society of Music Information Retrieval conference, 2017, Suzhou, P.R. China, pp. 575-581. Download the PDF

Scientific Publication: Emilia Parada-Cabaleiro, Maximilian Schmitt, Anton Batliner, Björn W. Schuller. Musical-Linguistic annotation of Il Lauro Secco. Proceedings of the 19th International Society of Music Information Retrieval conference, 2018, Paris, France, pp. 461-467. Download the PDF

Scientific Publication: Emilia Parada-Cabaleiro, Anton Batliner, Björn W. Schuller. A diplomatic edition of Il Lauro Secco: Ground truth for OMR of white mensural notation. Proceedings of the 20th International Society of Music Information Retrieval conference, 2019, Delft, The Netherlands, pp. 557-564. Download the PDF

Example:

Rebelo Dataset

Official websites: http://www.inescporto.pt/~arebelo/index.php and http://www.inescporto.pt/~jsc/projects/OMR/

Summary: Three datasets of perfect and scanned music symbols including an extensive set of synthetically modified images for staff-line detection and removal. Contains approximately 15000 music symbols.

Scientific Publication: A. Rebelo, G. Capela, and J. S. Cardoso, "Optical recognition of music symbols: A comparative study" in International Journal on Document Analysis and Recognition, vol. 13, no. 1, pp. 19-31, 2010. DOI: 10.1007/s10032-009-0100-1

Examples:

Remarks: The dataset is usually only available upon request, but with written permission of Ana Rebelo I hereby make the datasets available under a permissive CC-BY-SA license, which allows you to use it freely given you properly mention her work by citing the above mentioned publication: Download the dataset.

Fornes Dataset

Official website: http://www.cvc.uab.es/~afornes/

Summary: A dataset of 4100 black and white symbols of 7 different symbol classes: flat, natural, sharp, double-sharp, c-clef, g-clef, f-clef.

Scientific Publication: A.Fornés and J.Lladós and G. Sanchez, "Old Handwritten Musical Symbol Classification by a Dynamic Time Warping Based Method", in Graphics Recognition: Recent Advances and New Opportunities. Liu, W. and Lladós, J. and Ogier, J.M. editors, Lecture Notes in Computer Science, Volume 5046, Pages 51-60, Springer-Verlag Berlin, Heidelberg, 2008. DOI: 10.1007/978-3-540-88188-9_6

Example:

Choi Accidentals Dataset

Official website: https://www-intuidoc.irisa.fr/en/choi_accidentals/

Summary: A dataset of 2955 small black and white images of accidentals (flat, natural, sharp) in context, including 968 images without accidentals (reject class). Annotations are included into the filename such as {composer}-{page number}_{accidental class}_{window box}_{accidental box}_{note head box}.jpg with the boxes containing absolute coordinates, relative to the original music score page in the format: {left}x{top}x{right}x{bottom}.

Example:

Audiveris OMR

Official website: https://github.com/Audiveris/omr-dataset-tools

Summary: A collection of four music sheets with approximately 800 annotated music symbols. The DeepScore project in cooperation with the ZHAW targets towards automatically generating these images and the annotations from MuseScore or Lilypond documents.

Example:

Printed Music Symbols Dataset

Official website: https://github.com/apacha/PrintedMusicSymbolsDataset

Summary: A small dataset of about 200 printed music symbols out of 36 different classes. Partially with their context (staff-lines, other symbols) and partially isolated.

Example:

Music Score Classification Dataset

Official website: https://github.com/apacha/MusicScoreClassifier

Summary: A dataset of 2000 images, containing 1000 images of music scores and 1000 images of other objects including text documents. The images were taken with a smartphone camera from various angles and different lighting conditions.

Scientific Publication: Alexander Pacha, Horst Eidenberger, Towards Self-Learning Optical Music Recognition. 2017 16th IEEE International Conference on Machine Learning and Applications (ICMLA), Cancún, Mexiko, Dezember 2017. DOI: 10.1109/ICMLA.2017.00-60

Example:

OpenOMR Dataset

Official website: http://sourceforge.net/projects/openomr/

Summary: A dataset of 706 symbols (g-clef, f-clef) and symbol primitives (note-heads, stems with flags, beams) of 16 classes created by Arnaud F. Desaedeleer as part of his master thesis to train artificial neural networks.

Scientific Publication: Arnaud F. Desaedeleer, "Reading Sheet Music", Master Thesis, University of London, September 2006, Download

Example:

Gamera MusicStaves Toolkit

Official website: http://music-staves.sf.net/ and https://github.com/hsnr-gamera

Summary: The Synthetic Score Database by Christoph Dalitz that contains 32 scores that have been computer generated with different music typesetting programs. It contains ground truth data and is suitable for the deformations implemented in the toolkit.

Scientific Publication: C. Dalitz, M. Droettboom, B. Pranzas, I. Fujinaga: A Comparative Study of Staff Removal Algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 30, no. 5, pp. 753-766 (2008) DOI: 10.1109/TPAMI.2007.70749

Example:

Early Typographic Prints

Summary: 240 pages of early typographic music having a total of 1478 staves and 52178 characters corresponding to 175 different symbols with ground-truth obtained by manually entering via a MIDI keyboard.

Scientific Publication: Laurent Pugin. Optical Music Recognition of Early Typographic Prints using Hidden Markov Models. 7th International Conference on Music Information Retrieval (ISMIR’06), Victoria, Canada, October 2006. http://www.aruspix.net/publications/pugin06optical.pdf

Example:

Silva Online Handwritten Symbols

Summary: Dataset of 12600 trajectories of handwritten music symbols, drawn by 50 writers with an Android application. Every writer drew each of the 84 different symbols three times.

Scientific Publication: Rui Miguel Filipe da Silva. Mobile framework for recognition of musical characters. Master Thesis. Universidade do Porto, June 2013. https://repositorio-aberto.up.pt/bitstream/10216/68500/2/26777.pdf

IMSLP

Official website: http://imslp.org

Summary: The Petrucci Music Library is the largest collection of public domain music, with over 420000 (Jan. 2018) freely available PDF scores by almost 16000 composers accompanied by almost 50000 recordings. It also maintains an extensive list of other music score websites, where you can find many more music sheets, e.g. collected during research projects by universities.

Example:

Byrd Dataset

Official website: http://www.diku.dk/hjemmesider/ansatte/simonsen/suppmat/jnmr/ (broken). Download from Github mirror.

(Authors want to be contacted)

Summary: A small dataset of 34 high quality images with individual music score pages of increasing difficulty.

Scientific Publication: Donald Byrd & Jakob Grue Simonsen: "Towards a Standard Testbed for Optical Music Recognition: Definitions, Metrics, and Page Images". Journal of New Music Research, vol 44, nr.3, pages 169-195, 2015. DOI: 10.1080/09298215.2015.1045424

Example: