STORYBENCH
收藏资源简介:
STORYBENCH是由谷歌研究院创建的一个多面性基准数据集,用于连续故事可视化任务。该数据集包含来自三个现有视频数据集的6000个视频,通过收集密集、丰富的标注,包括动作描述、时间戳和多模态基础,以及每个视频段的标签,以便于确定失败模式。STORYBENCH旨在可靠地评估即将推出的文本到视频模型,并包含三个难度递增的视频生成任务:动作执行、故事延续和故事生成。数据集的目标是鼓励未来在真实世界、文本到视频生成方面的研究努力,并通过可重复和全面的设置激发更多工作。
STORYBENCH is a versatile benchmark dataset created by Google Research for sequential story visualization tasks. This dataset comprises 6000 videos sourced from three existing video datasets, with dense and rich annotations collected, including action descriptions, timestamps, multimodal grounding, and labels for each video segment, to facilitate the identification of failure modes. STORYBENCH aims to reliably evaluate upcoming text-to-video models, and includes three video generation tasks with increasing difficulty: action execution, story continuation, and story generation. The goal of this dataset is to encourage future research efforts on real-world text-to-video generation, and spur more work through a reproducible and comprehensive setup.




