VNBench
收藏资源简介:
VNBench是由中国科学院自动化研究所开发的视频理解综合基准数据集,旨在评估视频模型的细粒度理解和时空建模能力。该数据集通过合成视频生成,包含检索、排序和计数等任务,覆盖广泛的上下文长度。VNBench通过插入不相关的图像/文本'针'到原始视频中,生成仅基于这些'针'的标注,确保视频来源的多样性和查询-响应的多样性。此外,通过插入多个'针',VNBench严格评估模型的时序理解能力。该数据集的应用领域包括视频理解模型的评估和改进,特别是在处理长距离依赖任务方面。
VNBench is a comprehensive benchmark dataset for video understanding developed by the Institute of Automation, Chinese Academy of Sciences. It is designed to evaluate the fine-grained understanding and spatiotemporal modeling capabilities of video models. Generated via synthetic video production, this dataset encompasses tasks such as retrieval, ranking and counting, and spans a wide range of context lengths. VNBench inserts irrelevant image/text 'needles' into original videos to generate annotations solely based on these 'needles', ensuring the diversity of video sources and query-response pairs. Furthermore, by inserting multiple 'needles', VNBench rigorously assesses the temporal understanding capabilities of models. The application fields of this dataset include the evaluation and improvement of video understanding models, especially for tasks involving long-range dependencies.

- 1Needle In A Video Haystack: A Scalable Synthetic Framework for Benchmarking Video MLLMs中国科学院自动化研究所 · 2024年



