zzsi/deep-scores-v2

DeepScoresV2 — Complete

A HuggingFace-formatted mirror of the complete version of the DeepScoresV2 dataset for music object detection.

Dataset description

DeepScoresV2 is a large-scale dataset of synthetically rendered music score pages annotated with bounding boxes for musical symbols. The complete version contains 255,385 images with 151 million annotated instances across 135 symbol classes.

Each image is a full score page rendered from MuseScore across 5 music fonts (beethoven, emmentaler, gonville, gutenberg1939, lilyjazz). Annotations use COCO format: bbox is [x, y, width, height] in pixel coordinates.

Format

{
    "image_id":       int,
    "file_name":      str,
    "font":           str,           # music font used to render
    "image":          PIL.Image,
    "width":          int,
    "height":         int,
    "annotation_set": str,           # "deepscores" or "muscima++"
    "objects": {
        "id":          List[int],
        "bbox":        List[List[float]],  # [x, y, w, h], COCO format
        "category_id": List[int],
        "category":    List[str],
        "area":        List[float],
    },
}

Usage

from datasets import load_dataset

ds = load_dataset("zzsi/deep-scores-v2")
example = ds["train"][0]
print(example["objects"]["category"][:5])
example["image"].show()

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Attribution

This dataset is a reformatted mirror of DeepScoresV2. Please cite the original work:

@inproceedings{DeepScoresV2,
  title     = {DeepScoresV2: A Dataset for Music Object Detection with a Challenging Test Set},
  author    = {Tuggener, Lukas and Satyawan, Yvan Putra and Pacha, Alexander
               and Schmidhuber, J{\"u}rgen and Stadelmann, Thilo},
  booktitle = {British Machine Vision Conference (BMVC)},
  year      = {2021}
}

Original dataset: https://zenodo.org/records/4012193 Original authors: Lukas Tuggener, Yvan Putra Satyawan, Alexander Pacha, Jürgen Schmidhuber, Thilo Stadelmann (ZHAW / IDSIA)