Large-scale Cartoon Segmentation Dataset

本研究介绍了一个大规模的漫画分割数据集，包含98,600对高分辨率漫画图像及其实例标注掩码。数据集通过逆向工程方法创建，从色键视频和静态插图中提取角色和对象实例，并通过合成模拟漫画和动画的实际组成。此数据集旨在支持自动漫画编辑应用，如视觉风格编辑、运动分解和转移，以及为增强视觉体验计算立体深度。数据集的创建解决了现有分割方法在漫画领域中的局限性，提高了分割质量，适用于多种漫画编辑任务。

This study presents a large-scale comic segmentation dataset consisting of 98,600 pairs of high-resolution comic images and their corresponding instance annotation masks. The dataset is developed through reverse engineering: character and object instances are extracted from chroma-key videos and static illustrations, then synthesized to replicate the actual compositional structure of comics and animations. This dataset is designed to support automatic comic editing applications, including visual style editing, motion decomposition and transfer, as well as stereo depth calculation for enhancing visual experience. The creation of this dataset addresses the limitations of existing segmentation methods in the comic domain, improves segmentation quality, and is applicable to a variety of comic editing tasks.