BowenC/time-signature

Time Signature Detection Dataset

A YOLO-format object detection dataset for recognizing musical time signatures in sheet music images. The dataset is compiled from two sources: a subset of DeepScores V2 and public domain scores from IMSLP (Chopin, Beethoven, Mozart, Brahms, Debussy, and others).

Dataset Summary

Split	Images	Instances
Train	381	2,481
Val	269	644
Total	650	3,125

Classes

19 time signature classes with instance counts:

ID	Label	Meaning	Train	Val	Total
0	44	4/4	160	72	232
1	34	3/4	390	118	508
2	64	6/4	52	52	104
3	24	2/4	242	52	294
4	c	Common time	244	60	304
5	68	6/8	294	83	377
6	84	8/4	40	0	40
7	94	9/4	57	11	68
8	138	13/8	28	0	28
9	104	10/4	16	0	16
10	38	3/8	123	8	131
11	78	7/8	55	27	82
12	98	9/8	241	56	297
13	62	6/2	8	0	8
14	42	4/2	8	0	8
15	ch	Cut time (alla breve)	182	41	223
16	128	12/8	315	64	379
17	54	5/4	4	0	4
18	32	3/2	22	0	22
—	Total	—	2,481	644	3,125

Dataset Structure

dataset/
├── data.yaml
├── classes.txt
├── images/
│   ├── train/   # 381 PNG images
│   └── val/     # 269 PNG images
└── labels/
    ├── train/   # YOLO .txt annotations
    └── val/     # YOLO .txt annotations

Labels follow the YOLO format: <class_id> <x_center> <y_center> <width> <height> (normalized 0–1).

Usage

With Ultralytics YOLO

from ultralytics import YOLO

model = YOLO("yolov8n.pt")
model.train(data="data.yaml", epochs=100)

Download with huggingface_hub

from huggingface_hub import snapshot_download

path = snapshot_download(repo_id="BowenC/time-signature", repo_type="dataset")

Data Sources

This dataset is compiled from two sources:

• DeepScores V2 — Digitally rendered sheet music images with dense music symbol annotations. A subset of images was filtered and re-annotated for time signature detection.
• IMSLP (International Music Score Library Project) — Scanned public domain sheet music scores by composers including Chopin, Beethoven, Mozart, Brahms, and Debussy.

License

CC BY 4.0

DeepScores V2 is released under CC BY 4.0. IMSLP source scores are in the public domain. Annotations are original work by the dataset author.

Citation

If you use this dataset, please also cite the original sources:

DeepScores V2:

@dataset{deepscorev2_2020,
  author    = {Tuggener, Lukas and Elezi, Ismail and Schmidhuber, Jürgen and Pelillo, Marcello and Stadelmann, Thilo},
  title     = {{DeepScoresV2 Dataset and Benchmark for Music Object Detection}},
  year      = {2020},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.4012193},
  url       = {https://zenodo.org/records/4012193}
}

IMSLP:

IMSLP: Petrucci Music Library. https://imslp.org/