• DOI:10.48550/arXiv.2211.13285
• Corpus ID: 254017416

Proceedings of the 4th International Workshop on Reading Music Systems

@article{CalvoZaragoza2022ProceedingsOT,
  title={Proceedings of the 4th International Workshop on Reading Music Systems},
  author={Jorge Calvo-Zaragoza and Alexander Pacha and Elona Shatri},
  journal={ArXiv},
  year={2022},
  volume={abs/2301.10062}
}

• Jorge Calvo-Zaragoza, Alexander Pacha, E. Shatri
• Published 23 November 2022
• Computer Science
• ArXiv

—A major limitation of current Optical Music Recog- nition (OMR) systems is that their performance strongly depends on the variability in the input images. What for human readers seems almost trivial—e.g., reading music in a range of different font types in different contexts—can drastically reduce the output quality of OMR models. This paper introduces the 19MT-OMR corpus that can be used to test OMR models on a diverse set of sources. We illustrate this challenge by discussing several examples from this dataset. Abstract —There have been significant breakthroughs in com- puter vision research in many subfields, including composer identification from images of sheet music. Previous work in composer identification depends on a specific digital semantic representation of music and various evaluation criteria, making it difficult to quantify their relative merits. We present a novel approach using an end-to-end deep neural network model for music composer identification with images of sheet music as inputs. Hence, this method is not dependent on the conversion of the sheet music to any other intermediate digital semantic format. Additionally, we compare results from classification applied to sheet music and the respective bootleg representation. Identifying the composer can lead to more inferred data, which is helpful in archiving historical pieces digitally. Based on our experimental results, it can be concluded that the composer identification in sheet music images with deep neural models shows promising results. With the proposed model, we achieved 83% accuracy for composer identification on sheet music images compared to 76% accuracy when applied to the bootleg representations on our newly collected dataset. Abstract —Training with synthetic data has been successfully used in many domains of deep learning where authentic training data is scarce. Optical Music Recognition (OMR), especially recognition of handwritten music, greatly benefits from training on synthetic data too. In this paper, we explore the challenges of synthesizing images of sheets of music for training deep learning OMR models and compare such synthesis to the process of digital music engraving. We also contrast that with the architecture of our synthesizer prototype, which was used to achieve state-of- the-art results by training on the synthetic images only. Abstract —Full Page Optical Music Recognition (OMR) systems typically consist of multi-step workflows. However, the fine-tuning of these systems tends to be costly. We present the first layout analysis-free full-page OMR model that receives a page image and directly outputs its transcription in a single step. This model requires only the annotations of full score pages during training. The model has been tested with early-notation monophonic music scores, for which the presented approach is especially beneficial. Results show that this methodology provides a solution with promising results and establishes a new line of research for end-to-end music transcription. Abstract —Modern advances in computer reading technologies have found excellent results using end-to-end neural approaches. In addition to reducing the number of steps required, these formulations force the neural network to learn to take into account the contextual nature of visual languages to boost recognition. So far, however, such approaches are only feasible when the output can be represented as a sequence. This works for many old music notations, or even for modern monophonic scores, but it is insufficient for the highest degrees of complexity of music notation. Furthermore, music notation can be nicely represented using a graph structure. In this paper, we present a first attempt at an end-to-end approach to predict graphs from images, taking complex music-notation symbols as input. Abstract —Optical Music Recognition (OMR) is the field of research that studies how to computationally read music notation from written documents. Thanks to recent advances in computer vision and deep learning, there are successful approaches that are able to locate the notation elements out of a given image. However, the stage of notation assembly, in which these elements must be related to reconstructing the musical notation itself, has received little attention in the last years. Furthermore, given the large number of elements in a music score, this stage must be efficient enough to be useful in practice. In this work, a couple of neural approaches that perform this stage efficiently are studied. Our experiments using the MUSCIMA++ handwritten sheet music corpus show that there exists an underlying trade-off between effectiveness and efficiency, since each approach shows very good results in only one of these two aspects. We hope that this work represents a starting point for further research in this important stage of the OMR. Abstract —ConcertCue is a web application that synchronizes content delivery with live music. During a concert, a human operator supervises a score-following process to ensure it stays in sync. To aid the operator, the system displays a scanned copy of the score and highlights the current estimated measure, which requires the system to know the visual location of each measure on each page. In this paper, we present a fast, accurate OMR algorithm for identifying systems and barlines in a musical score and an integrated user interface that allows for quick OMR verification and correction. The use of this system as a part of ConcertCue has accumulated over 1,000 pages of human-verified annotations, which we plan to release as a dataset in the near future. Abstract —We present a new approach for the transcription of drum performances captured on a MIDI drum kit into scores in conventional Western notation. It works by parsing an input MIDI sequence into a tree-structured intermediate representation, using formal language techniques, post-processing using term rewriting, and finally exporting into an XML score file in the MEI encoding. An experimentation was conducted on the Groove MIDI Dataset. Abstract —Handwritten Music Recognition, especially in the historical domain, is an inherently challenging endeavour; paper degradation artefacts and the ambiguous nature of handwriting make recognising such scores an error-prone process, even for the current state-of-the-art Sequence to Sequence models. In this work we propose a way of reducing the production of statistically implausible output sequences by fusing a Language Model into a recognition Sequence to Sequence model. The idea is leveraging visually-conditioned and context-conditioned output distributions in order to automatically find and correct any mistakes that would otherwise break context significantly. We have found this approach to improve recognition results to 25.15 SER (%) from a previous best of 31.79 SER (%) in the literature.