DVQA Image Question Answering Dataset

This dataset is a research paper from Rochester Institute of Technology and Adobe Research DVQA: Understanding Data Visualizations via Question Answering The dataset proposed in . The dataset contains three folders: image folder, question-answer pair folder, and bar graph metadata folder.

Bar graphs are an effective way to convey numerical information, but today’s algorithms can’t parse them. Existing methods fail when faced with even small changes in appearance.The research team proposed DVQA, a dataset that tests many aspects of bar graph understanding in a question-answering framework.. Unlike visual question answering (VQA), DVQA needs to process words and answers that are unique to a particular bar graph. State-of-the-art VQA algorithms perform poorly on DVQA, and the research team proposed two strong baselines that perform better. The research team's work will enable algorithms to automatically extract numerical and semantic information from a large number of bar graphs in scientific publications, internet articles, business reports, and many other fields.

The DVQA dataset is generated using matplotlib. The dataset contains two types of questions: one is a general question that is common to every chart, and the other is a special question that is valid for a specific chart.

The questions cover three aspects: a) structural understanding, b) data retrieval, and c) data reasoning. See the example above.
The test set has two types of questions covering a range of words: Test-Familiar only includes words in the training set, and Test-Novel also includes newly appeared words.