SVTA (Synthetic Video-Text Anomaly benchmark)
收藏资源简介:
SVTA 数据集是由西安交通大学、合肥工业大学和澳门大学的研究团队创建的,旨在解决视频异常检索中的数据稀缺和隐私限制问题。该数据集包含了 41,315 个视频-文本对,涵盖了 68 种异常类型和 30 种正常活动,其中异常视频和正常视频的比例为 3:2。数据集采用了文本指导的视频生成模型,确保了视频内容与文本描述的一致性,同时避免了真实数据收集中的隐私问题。SVTA 数据集在视频异常检索领域具有重要的应用价值,有助于提高模型的鲁棒性和泛化能力。
The SVTA dataset was created by research teams from Xi'an Jiaotong University, Hefei University of Technology, and the University of Macau, aiming to address the issues of data scarcity and privacy constraints in video anomaly retrieval. This dataset contains 41,315 video-text pairs, covering 68 types of anomalies and 30 types of normal activities, with a 3:2 ratio between anomalous and normal videos. The dataset adopts text-guided video generation models, ensuring the consistency between video content and text descriptions, while avoiding privacy issues arising from real-world data collection. The SVTA dataset holds significant application value in the field of video anomaly retrieval, and helps improve the robustness and generalization ability of models.
- 1Towards Scalable Video Anomaly Retrieval: A Synthetic Video-Text Benchmark西安交通大学, 合肥工业大学, 澳门大学 · 2025年



