VERIFIED
收藏资源简介:
VERIFIED是由清华大学开发的自动视频文本标注管道,旨在生成具有可靠静态和动态细节的细粒度视频标注。该数据集包括Charades-FIG、DiDeMo-FIG和ActivityNet-FIG,这些数据集通过使用大型语言模型(LLM)和大型多模态模型(LMM)生成多样化的细粒度标注。数据集的创建过程结合了静态和动态增强的标注模块,并通过细粒度感知噪声评估器过滤不准确的标注。VERIFIED数据集主要应用于视频语料库时刻检索(VCMR)领域,旨在提高模型对视频中细粒度信息的理解和定位能力。
VERIFIED is an automatic video-text annotation pipeline developed by Tsinghua University, which aims to generate fine-grained video annotations with reliable static and dynamic details. This dataset includes Charades-FIG, DiDeMo-FIG, and ActivityNet-FIG, whose diverse fine-grained annotations are generated using Large Language Models (LLMs) and Large Multimodal Models (LMMs). The dataset creation process integrates static and dynamic enhanced annotation modules, and filters out inaccurate annotations via a fine-grained-aware noise evaluator. The VERIFIED dataset is primarily applied in the field of Video Corpus Moment Retrieval (VCMR), aiming to enhance models' ability to understand and localize fine-grained information in videos.




