ViTextVQA / README.md
nhonhoccode's picture
Update README.md
4d6a49a verified
metadata
dataset_info:
  viewer: false

📌 ViTextVQA — Vietnamese Text-based Visual Question Answering Dataset


🇻🇳 Giới thiệu (Tiếng Việt)

ViTextVQA là một dataset Visual Question Answering (VQA) dành cho tiếng Việt, tập trung vào khả năng đọc hiểu chữ xuất hiện trong ảnh (scene text), dựa trên bài báo ViTextVQA: A Large-Scale Visual Question Answering Dataset for Evaluating Vietnamese Text Comprehension in Images (ArXiv 2404.10652).
Bản release trên Hugging Face này đã được chỉnh sửa cấu trúc và bổ sung các file phục vụ việc huấn luyện mô hình.


🇺🇸 Introduction (English)

ViTextVQA is a Visual Question Answering (VQA) dataset for Vietnamese, focusing on scene text comprehension in images, as described in the paper ViTextVQA: A Large-Scale Visual Question Answering Dataset for Evaluating Vietnamese Text Comprehension in Images (ArXiv 2404.10652).
This Hugging Face release contains reorganized annotation format and additional supporting files for model training.


📁 Dataset Structure


ViTextVQA/
├── images.zip
├── train.json
├── dev.json
├── test.json
├── vitextvqa_coco.json
├── vitextvqa_captions.json
├── docr_features_of_vitext.npy
└── README.md
File Mô tả / Description
images.zip Ảnh dataset (đã nén) / All dataset images (zipped)
train.json COCO-like annotation cho split train
dev.json COCO-like annotation cho split validation
test.json COCO-like annotation cho split test
vitextvqa_coco.json (Optional) Original COCO-like annotations
vitextvqa_captions.json Caption/annotation bổ sung
docr_features_of_vitext.npy OCR / document features precomputed
README.md File mô tả này

🧠 Annotation Format (COCO-like)

Cấu trúc annotation ban đầu theo dạng COCO-like:

{
  "images": [
    { "id": 9836, "filename": "9836.jpg" },
    { "id": 14257, "filename": "14257.jpg" }
  ],
  "annotations": [
    {
      "id": 74,
      "image_id": 22,
      "question": "cửa tiệm màu xanh là gì ?",
      "answers": ["nhà thuốc"]
    },
    ...
  ]
}

Để training VQA hoặc dùng chung với frameworks như Hugging Face, bạn có thể convert mỗi annotation thành 1 sample:

{
  "image": "images/9836.jpg",
  "question": "cửa tiệm màu xanh là gì ?",
  "answers": ["nhà thuốc"],
  "question_id": 74
}

📊 Dataset Statistics / Thống kê

Split Images QA pairs (annotations)
train 11,733 35,159
dev 1,676 5,155
test 3,353 10,028

Các con số được tính bằng script thống kê annotation COCO-like.


🛠️ Usage / Hướng dẫn sử dụng

📦 Giải nén ảnh (unzip images)

unzip images.zip -d images/

📖 Load JSON với Python

import json

with open("train.json", "r", encoding="utf-8") as f:
    data = json.load(f)

print(data[0])

📚 Dùng dataset với Hugging Face datasets

from datasets import load_dataset

dataset = load_dataset("json", data_files="train.json")
print(dataset["train"][0])

📜 Citation / Trích dẫn

Nếu bạn sử dụng dataset này trong nghiên cứu, vui lòng trích dẫn:

@article{ViTextVQA2024,
  title={ViTextVQA: A Large-Scale Visual Question Answering Dataset for Evaluating Vietnamese Text Comprehension in Images},
  author={Quan Van Nguyen and Dan Quang Tran and Huy Quang Pham and Thang Kien-Bao Nguyen and Nghia Hieu Nguyen and Kiet Van Nguyen and Ngan Luu-Thuy Nguyen},
  journal={arXiv preprint arXiv:2404.10652},
  year={2024},
  url={https://arxiv.org/abs/2404.10652}
}

📬 Contact / Liên hệ

Dataset gốc được công bố bởi nhóm tác giả nghiên cứu tại University of Information Technology, Vietnam National University, Ho Chi Minh City.