task_categories:
- visual-question-answering
language:
- en
pretty_name: YTTB-VQA
size_categories:
- n<1K
license: cc-by-nc-4.0
Dataset Card for Dataset Name
Dataset Description
- Homepage: https://gordonhu608.github.io/bliva/
- Repository: https://github.com/mlpc-ucsd/BLIVA.git
- Paper:
- Point of Contact: w1hu@ucsd.edu
Dataset Summary
The YTTB-VQA Dataset is a collection of 400 Youtube thumbnail question-answer pairs to evaluate the visual perception abilities of in-text images. It covers 11 categories, including technology, sports, entertainment, food, news, history, music, nature, cars, and education.
Supported Tasks and Leaderboards
This dataset supports many tasks, including visual question answering, image captioning, etc.
License
CC-By-NC-4.0
Languages
The language of the data is primarily English.
Getting Started
Creating the dataset
Run the following command to download the images and create the dataset:
python3 create_dataset.py
You will find the images in images_new and the dataset in youtube_new.json.
Dataset Structure
Data Instances
A data instance in this dataset represents entries from a collection augmented by human-generated questions submitted to BLIVA. The answer is then entered into the answer field.
Data Fields
video_id: a unique string representing a specific YouTube thumbnail image.
question: representing a human-generated question.
video_classes: representing a specific category for the YouTube thumbnail image.
answers: This represents a ground truth answer for the question made about the YouTube thumbnail image.
video link Representing the URL link for each YouTube video.
Data Splits
The data are unsplit.
Dataset Creation
Source Data
Initial Data Collection and Normalization
We randomly selected YouTube videos with text-rich thumbnails from different categories during the data collection. We recorded the unique video ID for each YouTube video and obtained the high-resolution thumbnail from the URL ”http://img.youtube.com/vi/YouTube-Video-ID/maxresdefault.jpg”.
Annotations
Annotation process
We created the annotation file with the following fields: ”video id,” question,” video classes,” answers,” and ”video link" in JSON format.
Considerations for Using the Data
Discussion of Biases
Although our dataset spans 11 categories, the ratio within each category varies. For example, 18% of the dataset pertains to education, while only 2% is dedicated to news.
Acknowledgments
The youtube thumbnails dataset is purely for academic research and not for any monetary uses. For any of the authors who saw our dataset and found their thumbnail images used inappropriately, please get in touch with us directly by this email at w1hu@ucsd.edu and we will remove the image immediately.