Optical Music Recognition (OMR) has become a popular technology to retrieve information present in musical scores in conjunction with the increasing improvement of Deep Learning techniques, which represent the state-of-the-art in the field. However, its effectiveness is limited to cases where the target collection is similar in musical context and graphical appearance to the available training examples. To address this limitation, researchers have resorted to labeling examples for specific neural models, which is time-consuming and raises questions about usability. In this study, we propose a holistic and comprehensive study for dealing with new music collections in OMR, including extensive experiments to identify key aspects to have in mind that lead to better performance ratios. We resort to collections written in Mensural notation as specific use case, comprising 5 different corpora of training domains and up to 15 test collections. Our experiments report many interesting insights that will be important to create a manual of best practices when dealing with new collections in OMR systems.

If the video does not load properly please use the direct link to video