labos1/LSV

LSV是一个多视角的湿实验室生物学实验视频数据集，包含从第一人称（XMglass智能眼镜）、第三人称（DJI运动相机）和多视角（多个同步手机）拍摄的视频。每个视频记录了一名研究人员执行实验室协议的过程，并标注了相应的协议文本、场景类型以及（如适用）故意引入的程序错误。该数据集设计用于以下研究：协议合规性监控（检测程序是否正确执行）、程序错误检测（识别与标准协议的偏差）、自我中心视频理解（从第一人称视角理解实验室活动）、视频-语言基础（将协议文本与视频片段关联）以及多视角动作识别（利用同一程序的同时多视角）。数据集结构分为XMglass、DJI和Multiview三个子集，每个子集包含元数据、协议文本文件和视频文件。元数据字段包括唯一标识符、实验组标识符、录制日期、视频文件名、场景、操作描述、协议文件名、错误描述（如有）、视频时长、协议步骤时间戳、使用的实验室设备以及GPT-4o模型输出（如可用）。数据收集部分详细描述了每个子集的设备、条目数量、场景和注意事项。数据集涵盖了一系列常见的分子生物学和细胞培养技术，并包含许多故意引入的程序错误标注，用于自动化协议合规系统的基准测试。

LSV is a multi-view video dataset of wet-lab biology experiments, captured from a mix of first-person (XMglass smart glasses), third-person (DJI action camera), and multiview (multiple synchronized phones) perspectives. Each video records a researcher performing a laboratory protocol and is annotated with the corresponding protocol text, scene type, and—where applicable—deliberate procedural errors. The dataset is designed for research on protocol compliance monitoring, procedural error detection, egocentric video understanding, video-language grounding, and multiview action recognition. The dataset structure includes three subsets: XMglass, DJI, and Multiview, each containing metadata, protocol text files, and video files. Metadata fields include unique identifiers, experiment group identifiers, recording dates, video filenames, scenes, operation descriptions, protocol filenames, error descriptions (if present), video duration, protocol step timestamps, lab equipment used, and GPT-4o model outputs (where available). The data collection section details the devices, entry counts, scenes, and notes for each subset. The dataset covers a range of common molecular biology and cell culture techniques and includes annotations of deliberate procedural errors for benchmarking automated protocol-compliance systems.