A Unified Representation Framework for the Evaluation of Optical Music Recognition Systems

TL;DR

The need of a common music representation language is identified and the Music Tree Notation format is proposed, with the idea to construct a common endpoint for OMR research that allows coordination, reuse of technology and fair evaluation of community efforts.

Abstract

Modern-day Optical Music Recognition (OMR) is a fairly fragmented field. Most OMR approaches use datasets that are independent and incompatible between each other, making it difficult to both combine them and compare recognition systems built upon them. In this paper we identify the need of a common music representation language and propose the Music Tree Notation (MTN) format, with the idea to construct a common endpoint for OMR research that allows coordination, reuse of technology and fair evaluation of community efforts. This format represents music as a set of primitives that group together into higher-abstraction nodes, a compromise between the expression of fully graph-based and sequential notation formats. We have also developed a specific set of OMR metrics and a typeset score dataset as a proof of concept of this idea.

Figures (5)

• Figure 1: Fragment of a scan of a 1910 edition of Claude Debussy's La Danse de Puck from the first volume of his Préludes. In this fragment, many of the reasons why modelling OMR scores is difficult can be seen -- a two-staff system with voices that move along both of them, certain areas of the score accumulating many small symbols, many independent voices sounding in unison, elements spanning multiple measures, chords, etc.
• Figure 2: Example showing a fragment of a measure in which the annotation format for attributes and staff-modifying elements is shown. Rectangular nodes represent primitives as tokens and rounded nodes are abstract elements.
• Figure 3: The first few measures on Claude Debussy's first Arabesque. The single voice moves along both staves and is played using both hands on a piano. If one was to separate both staves, it would be impossible to extract the exact timing where each group of notes is to be played. Another interesting aspect of this piece is that many editions use implicit triplets, something that only can be inferred contextually by the fact that eighth notes are on groups of three.
• Figure 4: The pipeline through which the COMREF dataset has been generated.
• Figure 5: Some example measures sampled from the dataset. The dataset contains scores of varying types. On the left, an example of a pianoform score with three staves. In the middle, two examples of single-staff scores where the top one is part of a system of staves. On the right, a two-staff pianoform with two staves.