---
dataset_info:
  features:
  - name: image_id
    dtype: int64
  - name: image
    dtype: image
  - name: width
    dtype: int64
  - name: height
    dtype: int64
  - name: objects
    sequence:
    - name: id
      dtype: int64
    - name: area
      dtype: int64
    - name: bbox
      sequence: float32
      length: 4
    - name: category
      dtype:
        class_label:
          names:
            '0': Textline
            '1': Heading
            '2': Picture
            '3': Caption
            '4': Columns
  - name: ground_truth
    struct:
    - name: gt_parse
      struct:
      - name: headline
        sequence: string
      - name: textline
        sequence: string
  splits:
  - name: train
    num_bytes: 121950518885.056
    num_examples: 58738
  download_size: 134761522514
  dataset_size: 121950518885.056
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: mit
task_categories:
- text-classification
- visual-question-answering
- object-detection
language:
- dv
tags:
- dhivehi
- thaana
- ocr
- textline
pretty_name: od_page_annotations
size_categories:
- 10K<n<100K
---

# 📦 Dhivehi Synthetic Document Layout + Textline Dataset

This dataset contains **synthetically generated** image-document pairs with detailed layout annotations and ground-truth Dhivehi text extractions.  
It’s designed for document layout analysis , visual document understanding , OCR fine-tuning, and related tasks specifically for Dhivehi script.

## 📋 Dataset Summary

- **Total Examples**: ~58,738  
- **Image Content**: Synthetic Dhivehi documents generated to simulate real-world layouts, including headlines, textlines, pictures, and captions.  
- **Annotations**:
    - Bounding boxes (`bbox`)
    - Object areas (`area`)
    - Object categories (`category`)
    - Ground-truth parsed text, split into:
        - `headline` (major headings)
        - `textline` (paragraph or text body lines)

## ⚠️ Important Note

This dataset is **synthetic** — no real-world documents or personal data were used. It was generated programmatically to train and evaluate models under controlled conditions, without legal or ethical concerns tied to real-world data.

## 🏷️ Categories

| Label ID | Label Name  |
|----------|-------------|
| 0        | Textline    |
| 1        | Heading     |
| 2        | Picture     |
| 3        | Caption     |
| 4        | Columns     |

## 📁 Features

| Field         | Type                                     |
|--------------|-----------------------------------------|
| `image_id`   | int64                                    |
| `image`      | image                                    |
| `width`      | int64                                    |
| `height`     | int64                                    |
| `objects`    | List of:
  - `id`: int64
  - `area`: int64
  - `bbox`: [x, y, width, height] (float32)
  - `category`: label (class label 0–4) |
| `ground_truth.gt_parse` |
- `headline`: list of strings
- `textline`: list of strings |

## 📊 Split

| Split  | # Examples | Size (bytes)        |
|--------|------------|---------------------|
| Train  | 58,738     | ~121.95 GB          |

Also check the compressed version 📁 **Repository**: [Hugging Face Datasets](https://huggingface.co/datasets/alakxender/od-syn-page-annotations-com)

## 📦 Download

The dataset requires ~134.76 GB for download and ~121.95 GB storage after extraction.


## 🔧 Example Use (with 🤗 Datasets)

```python
from datasets import load_dataset

dataset = load_dataset("alakxender/od-syn-page-annotations")

categories = dataset.features["objects"].feature["category"].names
id2label = {i: name for i, name in enumerate(categories)}

print(id2label)

sample = dataset['train'][0]
print("Image ID:", sample['image_id'])
print("Image size:", sample['width'], "x", sample['height'])
print("First object category:", sample['objects']['category'][0])
print("First headline:", sample['ground_truth']['gt_parse']['headline'][0])
```

## 📊 Visualize

```python
import numpy as np
from PIL import Image, ImageDraw, ImageFont
from datasets import load_dataset

def get_color(idx):
    palette = [
        "red", "green", "blue", "orange", "purple", "cyan", "magenta", "yellow", "lime", "pink"
    ]
    return palette[idx % len(palette)]

def draw_bboxes(sample, id2label, save_path=None):
    """
    Draw bounding boxes and labels on a single dataset sample.

    Args:
        sample: A dataset example (dict) with 'image' and 'objects'.
        id2label: Mapping from category ID to label name.
        save_path: If provided, saves the image to this path.

    Returns:
        PIL Image with drawn bounding boxes.
    """
    image = sample["image"]
    annotations = sample["objects"]

    # Convert to PIL Image
    image = Image.fromarray(np.array(image))
    draw = ImageDraw.Draw(image)
    try:
        font = ImageFont.truetype("arial.ttf", 14)
    except:
        font = ImageFont.load_default()

    # Draw boxes and labels
    for category, box in zip(annotations["category"], annotations["bbox"]):
        x, y, w, h = box
        color = get_color(category)
        draw.rectangle((x, y, x + w, y + h), outline=color, width=2)
        label = id2label[category]
        bbox = font.getbbox(label)
        text_width = bbox[2] - bbox[0]
        text_height = bbox[3] - bbox[1]
        draw.rectangle([x, y, x + text_width + 4, y + text_height + 2], fill=color)
        draw.text((x + 2, y + 1), label, fill="black", font=font)

    if save_path:
        image.save(save_path)
        print(f"Saved image to {save_path}")
    else:
        image.show()

    return image

from datasets import load_dataset

# Load one sample
dataset = load_dataset("alakxender/od-syn-page-annotations", split="train[:1]")

# Get category mapping
categories = dataset.features["objects"].feature["category"].names
id2label = {i: name for i, name in enumerate(categories)}

# Draw bounding boxes on the first sample
draw_bboxes(
    sample=dataset[0],
    id2label=id2label,
    save_path="sample_0.png"
)
```