SoyVitou/Khmer-Text-Detection-0.2k

SoyVitou/Khmer-Text-Detection-0.2k

Khmer OCR dataset for scene text detection + transcription.

This dataset is packaged as a Hugging Face dataset using a single Parquet file:

• —train/metadata.parquet

✅ The image column is stored as embedded bytes inside the Parquet, so load_dataset() works without downloading a separate images folder.

Dataset format

Each row contains:

• —id (string): sample id
• —image (image): image object (decoded by datasets)
• —annotation (string): JSON-encoded list of text instances, each with:
• —coordinate: 8 integers [x1,y1,x2,y2,x3,y3,x4,y4] (quadrilateral polygon)
• —text: transcription string

Example annotation (decoded):

[
  {"coordinate": [76,146,453,146,453,199,76,199], "text": "..." },
  {"coordinate": [27,195,400,195,400,240,27,240], "text": "..." }
]

Load with 🤗 Datasets

from datasets import load_dataset
import json

ds = load_dataset("SoyVitou/Khmer-Text-Detection-0.2k", split="train")

ex = ds[0]
img = ex["image"]                 # PIL image
ann = json.loads(ex["annotation"])  # list of dicts

print(ex["id"])
print(img.size)
print(ann[0])

Notes

• —Coordinates are provided as a quadrilateral (4 points) in pixel space.
• —annotation is stored as JSON text to ensure maximum compatibility with the Hub viewer and datasets.

This dataset was prepared by Mr. Soy Vitou