Accidental detection dataset


This work takes place in the context of the thesis of of Kwon-Young Choi on the subject of resolving segmentation problems in dense and damaged printed piano scores from the 1750 to 1950 period.

This data set contains 2955 images with annotations for the task of detecting or rejecting the correct accidental for a given note head.
There are three classes of accidentals: natural (968 examples), sharp(777 examples) and flat (242 examples) with an additional reject class (968 examples).
The annotation can be found in the names of the images:
{composer}-{page number}_{accidental class}_{window box}_{accidental box}_{note head box}.jpg
The boxes are in absolute coordinates relative to the original music score page and are in the following format:


Download the data set as a zip file