3U-VQA

To tackle the challenge of obtaining out-of-distribution (OOD) data for LVQA models, we introduced a novel dataset named 3U-VQA dataset (Usual, Unusual and Unknown object scenarios for LVQA with difficulty scoring dataset). The dataset comprises question and image sets. Each instance in the questions set is associated with a set of features representing the question-related criteria set. The questions and their ground truth answers are written using placeholders for the objects and their features, which can be specified based on the user needs and requirements. When creating the questions, we avoided deliberately binary (Yes/No) questions to prevent potential bias caused by the question's type in the model response.