OmniDiff
收藏资源简介:
OmniDiff是一个包含324个复杂现实世界和3D合成场景的高质量数据集,涵盖了12种不同的变化类型,每种变化都伴有精细的人类注释,平均描述长度为60个词。数据集通过现场摄影、网络爬虫和3D渲染相结合的方式收集,既包括室内外各种日常环境的变化对,也利用Blender模拟真实世界的变化。数据集的构建分为两个主要阶段:变化对收集和差异描述收集,以确保数据集的全面性和准确性。该数据集旨在用于图像差异描述任务,以解决环境监测和监控系统中精确描述视觉差异的需求。
OmniDiff is a high-quality dataset comprising 324 complex real-world and 3D synthetic scenarios, covering 12 distinct types of changes. Each change is accompanied by fine-grained human annotations, with an average description length of 60 words. The dataset is collected through a hybrid workflow integrating on-site photography, web crawling, and 3D rendering. It includes both change pairs of various indoor and outdoor daily environments, and employs Blender to simulate real-world changes. The construction of the dataset consists of two primary stages: change pair collection and difference description collection, which ensures the dataset's comprehensiveness and accuracy. This dataset is intended for the image difference description task, to fulfill the requirement of accurately describing visual differences in environmental monitoring and surveillance systems.




