roaminsight/icdar-mini

ICDAR_mini Dataset

A balanced mini subset of the ICDAR (International Conference on Document Analysis and Recognition) dataset with 50 samples per language. Includes actual document images and ground truth OCR text.

Dataset Details

• Total Samples: 500
• Total Images: 500
• Languages: 10
  • Arabic (50 samples)
  • Bangla (50 samples)
  • Chinese (50 samples)
  • Hindi (50 samples)
  • Japanese (50 samples)
  • Korean (50 samples)
  • Latin (50 samples)
  • Mixed (50 samples)
  • None (50 samples)
  • Symbols (50 samples)

File Structure

├── README.md
├── icdar_mini_index.json          # Dataset metadata
├── icdar_mini_Arabic.json         # Language-specific data
├── icdar_mini_Bangla.json
├── ...
└── images/
    ├── tr_img_00001.jpg
    ├── tr_img_00002.jpg
    └── ... (500 image files)

Each language subset is stored as a separate JSON file, and corresponding images are in the images/ directory.

Data Format

Each sample is a row in the dataset with the following columns:

• image: Document image
• ground_truth: OCR ground truth text
• language: Language of the document (e.g., "Latin", "Arabic", etc.)
• sample_id: Unique identifier for the sample

Example sample:

{
  "sample_id": "icdar_tr_img_07848",
  "image_path": "images/tr_img_07848.jpg",
  "ground_truth": "Text content...",
  "metadata": {
    "dataset": "ICDAR",
    "language": "Latin",
    "num_text_lines": 49,
    "image_size": [3264, 2448]
  }
}

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("kenza-ily/icdar-mini")

# Access a sample
sample = dataset["train"][0]
print(f"Image: {sample['image']}")
print(f"Ground Truth: {sample['ground_truth']}")
print(f"Language: {sample['language']}")
print(f"Sample ID: {sample['sample_id']}")

# Iterate through samples by language
for sample in dataset["train"]:
    if sample['language'] == 'Latin':
        print(f"{sample['sample_id']}: {sample['ground_truth'][:50]}...")

Citation

Please cite the original ICDAR dataset if you use this subset in your research.

License

This subset follows the original ICDAR dataset license.