OVBench
收藏资源简介:
OVBench是一个专门为在线视频理解设计的问答基准,由南京大学、中国移动研究院和OpenGVLab等机构联合开发。该数据集包含5000个高质量的时空细节标注,涵盖了电影、教学、道路场景、户外、室内和开放领域等7个不同领域的数据。数据集的创建过程包括任务定义、数据收集、问答生成和多选题生成,确保了标注的高质量和多样性。OVBench旨在评估模型在在线视频流中的时空细节理解能力,适用于自动驾驶、机器人助手和监控系统等实时应用场景。
OVBench is a question-answering benchmark specifically designed for online video understanding, jointly developed by institutions including Nanjing University, China Mobile Research Institute, OpenGVLab, and others. This dataset contains 5,000 high-quality spatio-temporal detail annotations, covering data from 7 distinct domains including movies, teaching, road scenarios, outdoor, indoor, and open-domain scenarios. The dataset creation process covers task definition, data collection, question-answering generation and multiple-choice question generation, ensuring the high quality and diversity of annotations. OVBench aims to evaluate the spatio-temporal detail understanding capability of models in online video streams, and is suitable for real-time application scenarios such as autonomous driving, robotic assistants and monitoring systems.

- 1Online Video Understanding: A Comprehensive Benchmark and Memory-Augmented Method南京大学, 中国移动研究院, OpenGVLab, 上海人工智能实验室 · 2025年



