openpecha/OCR-Tibetan_line_to_text_benchmark

Tibetan OCR-line-to-text Benchmark Dataset

This repository hosts a line-to-text benchmark dataset to evaluate and compare Tibetan OCR models. The dataset includes diverse scripts, writing styles, and print methods, enabling comprehensive testing across multiple domains.

💽 Datasets Overview

• Features:
  • filename: Name of the file.
  • label: Ground truth text.
  • image_url: URL of the image.
  • BDRC_work_id: BDRC scan id for specific works.
  • char_len: Character count of the text.
  • script: Type of script (e.g., Uchen, Umed).
  • writing_style: Writing style (e.g., betsug, drutsa).
  • print_method: The printing method used.

Dataset Splits

The dataset is divided into several splits, each representing different sources of Tibetan texts:

• Norbuketaka:

  • Number of Examples: 223,080
  • Size: 92.4 MB

• Lithang_Kanjur:

  • Number of Examples: 41,784
  • Size: 30.1 MB

• Lhasa_Kanjur:

  • Number of Examples: 16,326
  • Size: 10.1 MB

• Betsug:

  • Number of Examples: 1,217
  • Size: 774 KB

• Google_Books:

  • Number of Examples: 75,168
  • Size: 26.6 MB

• Karmapa8:

  • Number of Examples: 1,317
  • Size: 1.0 MB

• Drutsa:

  • Number of Examples: 595
  • Size: 272 KB

• Derge_Tenjur:

  • Number of Examples: 61,656
  • Size: 50.6 MB

• NorbukatekaNumbers:

  • Number of Examples: 317
  • Size: 0.093 MB

• KhyentseWangpo:

  • Number of Examples: 143
  • Size: 0.0847 MB

Total Dataset Size

• Total Dataset Size: ~ 211.9 MB
• Total Examples: 496,634

🔨 Methodology

The benchmark dataset was curated from the test splits of a comprehensive Tibetan OCR model dataset. Each split was selected to include a diverse representation of:

• Scripts: Various Tibetan scripts such as Uchen, Drutsa, and more.
• Writing Styles: Handwritten and printed texts.
• Print Methods: Different printing techniques used.

This approach ensures the dataset reflects challenging and varied scenarios for OCR evaluation. The diversity across splits allows for detailed performance analysis of OCR systems, particularly in handling Tibetan scripts with different complexities.

📥 Usage

To load the dataset using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("openpecha/OCR-Tibetan_line_to_text_benchmark")