Semantic segmentation dataset of Chinese martial arts classic movie props(part)

Semantic segmentation is a hot area in computer vision (CV) research. The integration of computer vision and the film and television industry has put forward higher requirements for media data. At present, the analysis of film clips integrating high dynamics and continuous frames is one of the hot spots of video semantic segmentation. High-precision, high-volume data is the basis for proposing and improving deep learning models. However, the content labels of existing datasets are disorderly, which is difficult to uniformly evaluate and review, and the label category subjects are blurred, which directly affects the effectiveness of the model. Chinese martial arts films are a genre of films with Chinese national characteristics and the epitome of traditional Chinese culture. The "theme props" in martial arts films not only show various weapons in ancient China, but also integrate the values of "chivalry" and "righteousness" in traditional Chinese culture, becoming an indispensable part of martial arts movies. This study creatively performs artificial semantic segmentation of props in Chinese martial arts films, and establishes prop segmentation rules and general review standards. It includes frame pictures of movie clips at 4 frames per second, improved semi-automatic annotation of 6 props (weapons), and a visual version of pixel (mask) markers. As far as we know, this dataset is the largest semantic segmentation dataset on Chinese movie props to date. This dataset can be used for traditional inspection tasks in computer vision as well as related research on interdisciplinary and interdisciplinary techniques.

语义分割（Semantic Segmentation）是计算机视觉（Computer Vision，CV）领域的研究热点。计算机视觉与影视行业的融合对媒体数据提出了更高的要求。当前，融合高动态与连续帧的影视片段分析，是视频语义分割的研究热点之一。高精度、大规模的数据是提出并优化深度学习模型的基础。然而现有数据集的内容标注杂乱无章，难以开展统一评估与审核，且标签类别主题界定模糊，直接影响了模型的应用效果。
中国武侠电影是兼具鲜明民族特色与中华传统文化缩影的影视类型。武侠影片中的“主题道具”，不仅展现了中国古代各式兵器，更融入了中华传统文化中“侠”与“义”的价值理念，成为武侠电影不可或缺的组成部分。
本研究创新性地针对中国武侠电影中的道具开展人工语义分割工作，制定了道具分割规则与通用审核标准。该数据集包含以每秒4帧速率采样的影视片段帧画面、针对6类道具（兵器）的优化半自动标注，以及像素级（掩码）标注可视化成果。据我们所知，本数据集是目前为止规模最大的中国影视道具语义分割数据集。本数据集可应用于计算机视觉领域的传统检验任务，亦可支撑跨学科相关技术的研究工作。