Release with the pre-trained models, as published in the article A Baseline for General Music Object Detection with Deep Learning.

Also see the MusicObjectDetection repository for script to download and pre-process multiple datasets, as well as evaluate the final results.

This release includes the trained model that has been used for the first version of the full-page music object detector.

This tag contains the final version of the music object detector with cropped images from the MUSCIMA++ dataset, as presented at the DAS 2018 in Vienna.

Future work no longer uses small image-crops, but attempts to detect the music object in the full page to avoid issues related to cropping the image and stitching the results.

This release includes the trained model, that achieved the best results on the test sets with the full vocabulary of the MUSCIMA++ dataset, trained on images with staff lines: 81.56% mAP, 94.22% weighted mAP. If you want to test the model on your own image-crops, the MusicObjectDetector/demo-folder contains a self-contained script that takes a trained model and an input image and performs object detection on that image:

Simply replace the linked model from there, with this model and use a small, cropped image as input instead of the full page.