Arko007/assistive-ocr-data-acquisition

Assistive OCR Benchmark Data

Multilingual OCR benchmark for visually impaired assistance — Indian medicine labels, packaged goods, and signage in Bengali, Hindi, and English.

Dataset Summary

Property	Value
Total images	7,004 rows in manifest
Image sources	images/hf_medicines/, images/openfoodfacts/, images/synthetic/
Domains	medicine_packaging (6,588), packaged_goods (386), signage (30)
Languages	bn+en (5,337), hi+en (868), en (799)
Splits	dev (4,217), val (1,337), test (1,450)
License	CC-BY-SA-4.0 (OF images), CC-BY-SA-4.0 (synthetic), CC-BY-SA-4.0 (curated)

Data Sources

Source	Count	License
dmedhi/indian-medicines (HF)	6,552	Public HF dataset; images from 1mg.com product listings
Open Food Facts	362	CC-BY-SA-4.0
assistive-ocr-synthetic	90	CC-BY-SA-4.0

Manifest Schema (benchmark_manifest.csv)

Column	Description
id	Unique identifier
image_path	Relative path in images/ directory
transcript	Ground truth text
language_profile	Language composition (bn+en, hi+en, en)
domain	medicine_packaging, packaged_goods, signage
source_object_id	Original source identifier
split	dev, val, or test
annotation_status	Annotation quality level
source_dataset	Origin dataset name
source_license	License information

Directory Layout

.
├── benchmark_manifest.csv        # Full manifest (7,004 rows)
├── benchmark_manifest_summary.json
├── acquisition.log               # Pipeline execution log
├── errors.log                    # Error log
├── images/
│   ├── hf_medicines/             # 6,552 JPGs from dmedhi/indian-medicines
│   ├── openfoodfacts/            # 362 JPGs from Open Food Facts India
│   └── synthetic/                # 90 PNGs (PIL-generated)
└── thumbnails/                   # 45 sample thumbnails

Usage

from datasets import load_dataset

dataset = load_dataset("Arko007/assistive-ocr-data-acquisition", split="dev")
print(dataset[0]["image_path"], dataset[0]["transcript"])