PreVAD (Pre-training Video Anomaly Dataset)
收藏资源简介:
PreVAD数据集是由中国传媒大学研究团队构建的,是目前为止规模最大、多样性最丰富的视频异常数据集。它包含了35,279个视频,涵盖犯罪、交通、动物、事故和生产等多个领域,每个视频都带有详细的异常描述和多级类别标签。该数据集通过利用基础模型自动化数据清洗和注释的可扩展数据策展管道进行构建,显著降低了人工标注成本,同时确保了高质量。PreVAD的创建旨在增强模型在新的范式下的泛化能力,并为开放世界场景下的视频异常检测提供支持。
The PreVAD dataset, constructed by the research team from Communication University of China, is the largest and most diverse video anomaly dataset to date. It contains 35,279 videos spanning multiple domains including crime, traffic, animals, accidents, and industrial production, with each video accompanied by detailed anomaly descriptions and multi-level category labels. This dataset was built using a scalable data curation pipeline that leverages foundation models for automated data cleaning and annotation, which significantly reduces manual annotation costs while ensuring high data quality. The PreVAD dataset was developed to enhance the generalization ability of models under novel paradigms and support video anomaly detection in open-world scenarios.

- 1Language-guided Open-world Video Anomaly Detection中国传媒大学 · 2025年



