VIP

VIP数据集是由加州大学圣巴巴拉分校的研究团队开发，专注于视频链式思维的推理能力评估。该数据集包含1500个样本，涵盖了多个领域的真实视频，旨在通过视频填充和预测任务测试模型的多跳多帧视频推理能力。数据集通过自动方法提取关键帧，并提供两种文本表示：无结构的密集标题和结构化的场景描述（FAMOuS），以增强视频理解和生成任务的效率和鲁棒性。VIP数据集的应用领域包括视频理解和生成，旨在解决视频处理中的计算复杂性和推理挑战。

The VIP dataset was developed by a research team from the University of California, Santa Barbara, and focuses on evaluating video chain-of-thought reasoning capabilities. This dataset contains 1,500 samples covering real-world videos across multiple domains, aiming to test the multi-hop and multi-frame video reasoning abilities of models through video inpainting and prediction tasks. The dataset extracts key frames through automated methods, and provides two types of text representations: unstructured dense captions and structured scene descriptions (FAMOuS), to enhance the efficiency and robustness of video understanding and generation tasks. The application fields of the VIP dataset include video understanding and generation, and it aims to address the computational complexity and reasoning challenges in video processing.