VideoVista
收藏资源简介:
VideoVista是一个综合性的视频理解与推理评估基准,由哈尔滨工业大学深圳研究院创建。该数据集包含3,400个视频,涵盖14个类别,总时长从几秒到超过10分钟不等,共衍生出25,000个问题。数据集的创建过程利用了GPT-4模型和先进的视频分析工具,确保了数据的高质量和多样性。VideoVista主要用于评估大型多模态模型在视频理解任务中的表现,特别是解决视频内容分析中的细粒度任务和逻辑推理问题。
VideoVista is a comprehensive video understanding and reasoning evaluation benchmark developed by the Harbin Institute of Technology Shenzhen Research Institute. This dataset comprises 3,400 videos spanning 14 categories, with durations ranging from several seconds to over 10 minutes, and a total of 25,000 derived questions. The dataset construction process leverages GPT-4 models and advanced video analysis tools to ensure high data quality and diversity. VideoVista is primarily used to evaluate the performance of large multimodal models in video understanding tasks, particularly for addressing fine-grained tasks and logical reasoning problems in video content analysis.

- 1VideoVista: A Versatile Benchmark for Video Understanding and Reasoning哈尔滨工业大学深圳研究院 · 2024年



