Fully Authentic Visual Question Answering Dataset from Online Communities

University of Texas at Austin Microsoft Research University of Colorado Boulder,
ECCV 2024

VQAonline

Overview

VQAonline is the first VQA dataset in which all contents originate from an authentic use case. It is sourced from an online question answering community (i.e., StackExchange). It differs from prior datasets; examples include that it contains: (1) authentic context that clarifies the question, (2) an answer the individual asking the question validated as acceptable from all community provided answers, (3) answers that are considerably longer (e.g., a mean of 173 words versus typically 11 words or fewer in prior work), and (4) user-chosen topics for each visual question from 105 diverse topics revealing the dataset’s inherent diversity.

The VQAonline dataset introduces new challenges to current vision and language foundation models (VLMs). These challenges include: (1) Visual questions (VQs) that demand advanced reasoning abilities, such as those found in "puzzling" topics. (2) VQs that require diverse types of knowledge, spanning everyday-life knowledge, fine-grained knowledge, and expert knowledge. (3) Novel image types such as 3D renderings and sheet music, which are the most challenging image types for current VLMs. (4) Multilingual VQs in VQAonline highlights the varying performance of current VLMs across different languages.

VQAonline Dataset

The VQAonline dataset includes:


Training, Validation, and Test Sets

  • Training Set
    • 665 examples
  • Validation Set
    • 285 examples
  • Test Set
    • 63,746 examples

Details about each visual question are in the following format:

"image":"VQAonline_00021900.png",
"question":"What is the history of this coin? By Mandate of Heaven) Emperor of the Great Qin", Nurhaci - 1616-1626", 
"context":"I want to know the complete history of this coin. ",
"answer":"This particular coin is part of a set of commemorative tokens (aka fantasy coins) made in (modern) China, with a token for each of the Qing Dynasty emperors. This one shows Emperor Nurhaci with the dates when he was in power shown below his likeness on the "coin".",
"topic":"history",
"url":"https://history.stackexchange.com/questions/26573/"

compared to existing datasets

Characterization of existing VQA datasets and our VQAonline dataset in terms of mean question length (i.e., Q), mean answer length (i.e., A), number of images (i.e., Nimg), number of authentic topics (i.e., # Auth. Topics), inclusion of context, inclusion of authentic visual questions (i.e., Auth. VQ), inclusion of authentic context, and inclusion of answers validated by those asking the questions (i.e., Vld. A)).

Data Analysis

Algorithm Analysis

Error Examples

BibTeX

@article{chen2023vqaonline,
        title={Fully Authentic Visual Question Answering Dataset from Online Communities}, 
        author={Chen, Chongyan and Liu, Mengchen and Codella, Noel and Li, Yunsheng and Yuan, Lu and Gurari, Danna},
        journal={arXiv preprint arXiv:2311.15562},
        year={2023}}
@article{liu2023evalvqaonline,
        title={An Evaluation of GPT-4V and Gemini in Online VQA},
        author={Liu, Mengchen and Chen, Chongyan and Gurari, Danna},
        journal={arXiv preprint arXiv:2312.10637},
        year={2023}}

Contact

For any questions, comments, or feedback, please send them to Chongyan Chen at chongyanchen_hci@utexas.edu.