Dataset created by attoPascal for Optical Music Recognition. You can download the original version in his webpage.
The following changes from the original dataset have been done due to wrong labeling:
note-eighth-c1-3272 -> note-eighth-c2-3272
note-half-other-1223 -> note-half-g-1223
rest-half-2966 -> rest-whole-2966
rest-half-2991 -> rest-whole-2991
rest-half-3217 -> rest-whole-3217
rest-half-3999 -> rest-whole-3999
rest-half-4216 -> rest-whole-4216
rest-half-4296 -> rest-whole-4296
rest-half-4397 -> rest-whole-4397
rest-whole-4508 -> rest-half-4508
Distribution of the dataset according to the duration of each figure (The class 'other' are mostly triplets):
| whole | half | quarter | eighth | sixteenth | other | |
|---|---|---|---|---|---|---|
| note | 49 | 449 | 1184 | 1304 | 276 | 10 |
| rest | 43 | 26 | 216 | 266 | 1 | 0 |
In the next heatmap we can see what classes are poorly represented:
If you don't want difficult examples, you can delete the following images:
note-eighth-h1-4158
note-eighth-h1-4493
note-half-g1-3599
