VizWiz Visual Question Answering Dataset for the Blind

VizWiz-VQA (Visual Question Answering) is an image dataset for visual question answering for the blind. Blind users use the VizWiz software to take a photo and record a verbal question about the photo and 10 crowdsourced answers to the question. This dataset is used to solve the following two problems: one is to predict the answer to a visual question, and the other is to determine whether a visual question can be answered. This dataset aims to study more general algorithms to help blind people solve life obstacles.

The dataset includes (2020 latest version):

• 20,523 pairs of training images/questions
• 205,230 for training answers/answer confidence
• 4319 Verification images/questions
• 43,190 Verification answers/answer confidence
• 8,000 pairs of test images/questions