VARSew: A Visual Action Recognition Dataset on Garment Sewing

VARSew dataset is designed for visual action recognition in the garment sewing industry. Also, VARSew is designed to address multiple research objectives in visual recognition, such as binary and multi-class action classification. The proposed data offers new research opportunities, such as visual pattern recognition and Internet-of-Things(IOT) based manufacturing. Also, it offers research opportunities on human activity and behavior during production time. To the best of our knowledge, no such industrial human action dataset is available in the computer vision community.
VARSew dataset consists of high-resolution trimmed videos of certain actions. There are two levels of categorical labels in the VARSew dataset, i.e., super action classes and actions classes. The dataset includes 3,121 videos of 49,936 frames.
The videos are grouped into value-added and non-value-added for human sewing action binary classification. For multi-class classification, the videos are also grouped into eight classes: sew, release, handle, prepare, adjust, wait, check, and maintain. Each video was fixed to 16 frames. Multiple human operators were employed to collect the videos, including Sewing Machine Operators (SMO) and Maintenance Machine Operators (MMO), respectively 26 and 8 operators. Subjective annotations for the participant proficiency throughout the 3,121 videos.

VARSew数据集专为服装缝纫行业的视觉动作识别任务打造，同时可用于解决视觉识别领域的多项研究目标，涵盖二元动作分类与多分类动作分类任务。本数据集为相关研究提供了全新契机，包括视觉模式识别、基于物联网（Internet-of-Things, IoT）的智能制造方向，同时也为探究生产过程中的人类活动与行为提供了研究空间。据我们所知，当前计算机视觉领域尚未公开同类工业人类动作数据集。
VARSew数据集包含特定缝纫动作的高分辨率剪辑视频，设有两级分类标签体系，分别为动作超类与动作细类。数据集共计包含3121段视频，总帧数达49936帧。
针对人类缝纫动作的二元分类任务，视频被划分为增值类与非增值类两类；针对多分类任务，视频则被划分为8个细分类别：缝纫（sew）、释放（release）、操作（handle）、准备（prepare）、调整（adjust）、等待（wait）、检查（check）与维护（maintain）。每段视频均统一采样为16帧。
本次数据采集共聘请多名操作员参与，其中缝纫机操作员（Sewing Machine Operators, SMO）26名、设备维护操作员（Maintenance Machine Operators, MMO）8名。我们还针对全部3121段视频，对参与者的操作熟练度进行了主观标注。