CinePile
收藏资源简介:
CinePile是马里兰大学和魏茨曼科学研究所联合构建的一个针对长视频理解而设计的大型数据集。该数据集包含约305,000个多项选择题,源自9396个视频片段,涵盖了对视频内容的多维度理解,包括时间理解、人物与物体交互以及场景内事件或动作的推理等方面。数据集的构建过程采用自动化问题生成与验证的方法,结合了大模型和人工审核,确保了问题的质量和多样性。CinePile不仅为视频理解模型提供了一个全面的评估基准,也可用于视频内容分析和智能视频推荐,特别强调长视频的连贯性和上下文理解。
CinePile is a large-scale dataset designed for long-form video understanding, jointly constructed by the University of Maryland and the Weizmann Institute of Science. It contains approximately 305,000 multiple-choice questions derived from 9,396 video clips, covering multi-dimensional comprehension of video content, including temporal comprehension, human-object interactions, and reasoning about intra-scene events or actions, among other aspects. The dataset was built using an automated question generation and validation pipeline that integrates large language models (LLMs) and human review, ensuring the quality and diversity of the questions. CinePile not only serves as a comprehensive evaluation benchmark for video understanding models, but also can be applied to video content analysis and intelligent video recommendation, with a particular focus on long-form video coherence and contextual comprehension.




