Iftesha/odia-ocr-benchmark

Odia OCR Benchmark Dataset

Description

A curated benchmark dataset for evaluating OCR models on Odia (Oriya) text recognition. Contains handwritten, printed, scene text, newspaper, books, and digital categories, including both short samples and long-text examples for OCR evaluation.

Dataset Structure

• image: The input image (PIL Image)
• ground_truth: The correct Odia text transcription
• category: Type of text (handwritten, printed, scene_text, newspaper, books, digital)
• text_length: short (1-3 words), medium (4-10 words), long (10+ words)
• source: Original dataset source or "manual"

Usage

from datasets import load_dataset

dataset = load_dataset("Iftesha/odia-ocr-benchmark")

# Access a sample
sample = dataset["train"][0]
image = sample["image"]
text = sample["ground_truth"]

Categories

Category	Description
handwritten	Handwritten Odia text (word/short phrase level)
printed	Printed/typed Odia text
scene_text	Text in natural scenes (signboards, posters, etc.)
newspaper	Odia newspaper clippings (including long text)
books	Scanned Odia book pages (including long text)
digital	Screenshots from Odia digital content

Sources

• OdiaGenAIOCR/odia-ocr-merged (handwritten)
• darknight054/indic-mozhi-ocr with config oriya (printed)
• darknight054/indicstr12-crops with config odia (scene_text)
• newspaper: Odia newspaper scans/clippings
• books: Odia book page images
• digital: odia digital content

Notes

• Includes long-text samples for paragraph-level OCR evaluation.
• The source field records origin for each sample.

License

CC-BY-4.0