This project was about using the DETR architecture to identify and classify musical objects and their relations in images of handwritten sheet music (in western music notation). The dataset used can be found here: MUSCIMA++. Our main contribution was extending the DETR architecture to be able to classify the relations between objects such that end-to-end object detection and notation assembly can be handled holistically.
This repository contains the code used to train and evaluate our model. The training was partly done on a AAU provided Nvidia DGX System and on Google Cloud using the AI Platform. The thesis report is located in the repository as well mi1014f21_thesis.pdf.