VQA-VS

VQA-VS是由中国科学院信息工程研究所创建的一个视觉问答数据集，旨在评估模型在视觉问答任务中的推理能力，特别是在克服数据集偏差方面的能力。该数据集包含164,264条记录，通过引入多种类型的分布偏移，构建了多个OOD测试集，以确保模型不能仅依赖于训练集中的快捷解决方案。VQA-VS数据集的应用领域主要集中在提升视觉问答模型的泛化能力，解决模型在面对未知或多样化的数据分布时的表现问题。

VQA-VS is a visual question answering (VQA) dataset developed by the Institute of Information Engineering, Chinese Academy of Sciences. It is designed to evaluate the reasoning capabilities of models in VQA tasks, especially their ability to mitigate dataset biases. The dataset contains 164,264 records, and multiple out-of-distribution (OOD) test sets are constructed by introducing various types of distribution shifts, ensuring that models cannot rely solely on shortcut solutions from the training dataset. The main application areas of the VQA-VS dataset focus on enhancing the generalization ability of VQA models and addressing the performance issues of models when facing unknown or diverse data distributions.