CLEVR-Dialog

CLEVR-Dialog是由卡内基梅隆大学开发的一个大型诊断数据集，专注于研究视觉对话中的多轮推理。该数据集基于CLEVR图像，通过构建对话语法，为约85,000张图像生成了5个10轮对话实例，总计425,000个问答对。CLEVR-Dialog的特点是所有视觉对话方面均完全标注，适用于研究视觉参照解析等复杂挑战。数据集的应用领域包括训练和评估视觉对话模型，特别是在处理视觉参照和多轮对话推理方面的能力。

CLEVR-Dialog is a large-scale diagnostic dataset developed by Carnegie Mellon University, focusing on multi-turn reasoning in visual dialogue. By constructing dialogue grammars, it generates 5 instances of 10-turn dialogues for approximately 85,000 images, resulting in a total of 425,000 question-answer pairs. A notable feature of CLEVR-Dialog is that all aspects of visual dialogue are fully annotated, making it suitable for studying complex challenges such as visual reference resolution. The dataset can be employed to train and evaluate visual dialogue models, especially their capabilities in handling visual reference and multi-turn dialogue reasoning.