VANE-Bench

VANE-Bench是由穆罕默德·本·扎耶德人工智能大学创建的视频异常评估基准，包含325个视频片段和559个挑战性的问答对。该数据集结合了真实世界监控视频和AI生成的视频，涵盖多种异常类型，如不自然的变换、外观、穿透、消失和突然出现。创建过程中，通过半自动化的流程，包括帧标注、标题生成和问答生成模块，确保数据集的高质量和挑战性。VANE-Bench旨在评估和提升大型多模态模型在视频异常检测领域的应用，解决如深度伪造识别、视频内容操纵检测等实际问题。

VANE-Bench is a video anomaly evaluation benchmark developed by Mohamed bin Zayed University of Artificial Intelligence. It contains 325 video clips and 559 challenging question-answer pairs. This dataset combines real-world surveillance videos and AI-generated videos, covering multiple anomaly types including unnatural transformations, appearance anomalies, penetration, disappearance, and sudden emergence. During its construction, a semi-automated pipeline incorporating frame annotation, caption generation, and question-answer generation modules was utilized to guarantee the high quality and challenging nature of the dataset. VANE-Bench is designed to evaluate and improve the deployment of large multimodal models in the domain of video anomaly detection, and to resolve practical problems such as deepfake recognition and video content manipulation detection.