Rotations Are All You Need: A Generic Method For End-To-End Optical Music Recognition

Abstract

End-to-end Optical Music Recognition traditionally involves multi-step processes to address complex documents, where single stave isolation is performed. These state-of-the art methods often fall short in handling diverse music textures, such as polyphony or Aligned Music Notation and Lyrics Transcription (AMNLT). We introduce a generic end-to-end OMR approach compatible with monophonic, polyphonic, and AMNLT systems. By leveraging score rotations and multi-line transcription unfolding, this model only requires input system annotations during training. Experimental evidence suggests this approach offers a competitive solution with encouraging outcomes, paving the way for future end-to-end music transcription research.