VinaBench
收藏资源简介:
VinaBench数据集是由瑞士洛桑联邦理工学院、日本索尼集团和美国卡内基梅隆大学共同构建的视觉-文本叙事对数据集。该数据集包含约25K对从不同视觉叙事数据集中采样的视觉和文本叙事,并加入了常识和话语约束,以促进视觉叙事生成的一致性和忠实性学习。数据集详细标注了视觉实体的常识链接和话语特征,为视觉叙事生成模型的训练和评估提供了系统性的支撑。
The VinaBench dataset is a visual-text narrative pair dataset jointly constructed by École Polytechnique Fédérale de Lausanne (EPFL), Sony Group Corporation of Japan, and Carnegie Mellon University (CMU) of the United States. This dataset contains approximately 25,000 pairs of visual and textual narratives sampled from diverse visual narrative datasets, and incorporates common sense and discourse constraints to facilitate the learning of consistency and faithfulness in visual narrative generation. The dataset comprehensively annotates common sense links and discourse features of visual entities, providing systematic support for the training and evaluation of visual narrative generation models.




