Eagle-Video-110K
收藏资源简介:
Eagle-Video-110K是一个专门设计用于增强长视频理解能力的数据集,由南京大学等机构创建。该数据集整合了故事级别和剪辑级别注释,能够促进长视频的理解。数据集通过多样性驱动的方法收集,使用多个视频源和一个相似性阈值方法来识别新颖的片段,以最大化内容的多样性。数据集采用自上而下的故事级别方法和自下而上的剪辑级别方法进行注释,形成了密集的字幕,为全面的长形式问答对捕捉整个视频的叙事结构提供了基础。
Eagle-Video-110K is a dataset specifically tailored for advancing long-form video understanding, developed by institutions including Nanjing University. This dataset integrates both story-level and clip-level annotations, which enables improved comprehension of long videos. Collected via a diversity-driven methodology, it leverages multiple video sources and a similarity threshold technique to identify novel segments, thus maximizing content diversity. Annotations are generated using both top-down story-level and bottom-up clip-level approaches, yielding dense captions that establish a solid foundation for capturing the narrative structure of entire videos to support comprehensive long-form question-answering pairs.

- 1Eagle 2.5: Boosting Long-Context Post-Training for Frontier Vision-Language Models南京大学, 香港理工大学, 罗格斯大学 · 2025年



