CinePile
收藏资源简介:
CinePile是一个专为长视频理解设计的数据集,由马里兰大学帕克分校和魏茨曼科学研究所创建。该数据集包含305,000个多选题(MCQs),覆盖视觉和多模态方面,包括时间理解、人-物交互理解及场景内事件或动作的推理。数据集的创建过程结合了先进的LLMs和人工干预,利用现有的音频描述数据,并与YouTube上的公开电影视频片段对齐。CinePile的应用领域主要集中在视频理解,旨在解决现有数据集在长视频理解方面的不足,提供一个全面的模型性能评估基准。
CinePile is a dataset specifically designed for long-form video understanding, developed by the University of Maryland, College Park and the Weizmann Institute of Science. This dataset contains 305,000 multiple-choice questions (MCQs) covering visual and multimodal aspects, including temporal comprehension, human-object interaction understanding, and reasoning about in-scene events or actions. The construction of CinePile integrates state-of-the-art large language models (LLMs) and human intervention, leveraging existing audio description data and aligning with publicly available movie video clips on YouTube. The primary application scenarios of CinePile focus on video understanding, aiming to address the limitations of existing datasets in long-form video understanding and provide a comprehensive benchmark for model performance evaluation.

- 1CinePile: A Long Video Question Answering Dataset and Benchmark马里兰大学帕克分校 · 2024年



