Experiments on linking the nodes of a music notation graph (MuNG). The basic idea is to machine-learn the relations from images between objects in a notation graph. So starting with the detected object in an image, we want the machine to learn how to construct a Music Notationg Graph like this:
One way how this can be done is by training a convolutional neural network to take the output of an object detector and for each pair of objects decide whether an edge between them should be constructed or not.
More information can be found in the ISMIR paper.
- Python 3.6 / 3.7
- PyTorch
Install the sources with pip install -e . to follow updates without having to re-install.
Download the MUSCIMA++ dataset with the required images, by running
python munglinker/prepare_dataset.py
or manually download the MUSCIMA++ dataset and extract the xml-files from data/cropobjects_withstaff into data/mungs as well as the used images from the CVC-MUSCIMA dataset from here into data/images
Run the training by calling:
python munglinker/train.py