# YOLO Dataset — Flat & Alto Clef Detection

A YOLO-format object detection dataset for recognising two music notation symbols on orchestral score images: **flat accidentals** and **alto clefs**.

## Classes

| ID | Name | Description |
|----|------|-------------|
| `0` | `flat` | Flat accidental (♭) appearing in key signatures or as an accidental before a note |
| `1` | `alto_clef` | Alto (C) clef, typically used for the viola staff |

## Dataset Statistics

| Split | Images | `flat` | `alto_clef` |
|-------|--------|--------|-------------|
| train | 86 | 5390 | 411 |
| val | 20 | 1322 | 105 |
| test | 17 | 697 | 78 |
| **total** | **123** | **7409** | **594** |

## Source Scores

Images are cropped staff lines drawn from 7 orchestral scores:

| Score | Composer |
|-------|----------|
| beethoven7, beethoven8, beethoven10, beethoven11 | Ludwig van Beethoven |
| mendelssohn1 | Felix Mendelssohn |
| tchaikovsky0, tchaikovsky2 | Pyotr Ilyich Tchaikovsky |

## Structure

```
dataset_yolo_flat_alto/
├── dataset.yaml            # YOLO dataset config
├── images/
│   ├── train/              # 86 images
│   ├── val/                # 20 images
│   └── test/               # 17 images
└── labels/
    ├── train/              # YOLO .txt annotations
    ├── val/
    └── test/
```

## Image Naming

```
{score_name}_{staff_index}.jpg
```

Each image is a single cropped staff line extracted from the source score. For example, `beethoven10_3.jpg` is the 3rd staff crop from the score `beethoven10`.

## Label Format

Labels follow the standard YOLO format — one `.txt` file per image, one bounding box per line:

```
<class_id> <x_center> <y_center> <width> <height>
```

All values are normalised to `[0, 1]` relative to the image dimensions.

Example:
```
0 0.417922 0.213076 0.006325 0.011066
1 0.081175 0.877258 0.017169 0.018744
```

## YOLO Config

`dataset.yaml`:
```yaml
train: images/train
val:   images/val
test:  images/test
nc: 2
names: ['flat', 'alto_clef']
```