PaSBench-Video
收藏资源简介:
PaSBench-Video是由香港中文大学·深圳与清华大学联合创建的前瞻性视频安全预警评测基准,旨在评估多模态大语言模型在连续视频流中进行实时风险识别与预警的能力。该数据集包含740个视频样本,涵盖驾驶、医疗保健、日常生活及工业生产四大领域,每个风险视频均标注了帧级风险起始点与事故发生点,以支持时序精准度评估。数据通过融合公开真实视频与合成渲染技术构建,并经过严格的模型与人工双重筛选流程,确保风险的可预测性与视觉可辨识性。该数据集主要应用于主动安全监控系统研发,旨在解决现有模型在动态场景中因依赖表面线索而导致的误报率高与预警时机不准等核心问题。
PaSBench-Video is a prospective video safety warning evaluation benchmark jointly developed by The Chinese University of Hong Kong, Shenzhen and Tsinghua University, aiming to assess the capabilities of multimodal large language models (LLMs) to conduct real-time risk identification and early warning in continuous video streams. This dataset contains 740 video samples covering four core domains: driving, healthcare, daily life, and industrial production. Each risk video is annotated with frame-level risk start points and accident occurrence points to support temporal precision evaluation. The dataset is constructed by integrating publicly available real-world videos and synthetic rendering technologies, and has undergone strict dual screening processes including model-based validation and manual review, to ensure the predictability and visual recognizability of the included risks. This dataset is primarily utilized for the development of active safety monitoring systems, with the goal of addressing key challenges faced by existing models in dynamic scenarios, such as high false alarm rates and imprecise warning timing caused by over-reliance on superficial cues.
PaSBench-Video 数据集概述
PaSBench-Video 是一个用于主动安全与风险预判的视频基准数据集。每个样本包含一段视频及其对应的帧/时间标注和经过审核的风险响应提示。
数据集结构
metadata.csv:Hugging Face VideoFolder 元数据,file_name列指向视频文件。videos/:481 个精选视频文件,按原始来源数据集分组。annotations/final_annotations.json:经过清洗的公开标注文件,用于生成metadata.csv。annotations/label_schema.json:字段说明与数值分布。annotations/excluded_from_legacy_495.txt:旧版 495 文件包中存在但最终 481 版中排除的视频列表。
数据规模
- 视频总数:481
- 总字节数:4,064,785,884
- 各来源视频数量:
- DADA:105
- healthbench:33
- industry:46
- nexar:96
- oops:71
- radsv:17
- smarthome:113
字段说明
sample_id:稳定的 PaSBench 样本标识符。dataset:原始来源数据集/领域。risk_start_time_sec:预判风险开始时间(秒)。accident_time_sec:不良事件或损失开始时间(秒)。risk_reason:经审核的风险原因文本(优先使用修正版,否则使用原始文本)。recommended_user_action:经审核的建议行动文本(优先使用修正版,否则使用原始文本)。file_name:指向视频文件的 VideoFolder 技术键。
注意事项
最终精选集包含 481 条记录。旧版包描述中提及的 495 个视频属于遗留元数据;本次上传遵循最终审核的 481 版标注文件。
许可证与标签
- 许可证:cc-by-nc-4.0
- 标签:
video、safety、risk-anticipation、accident-anticipation





