fmow-fake-small
收藏资源简介:
fmow-fake-small是由橡树岭国家实验室构建的卫星图像操纵与深度伪造定位基准数据集,旨在弥补现有遥感伪造数据集缺乏高质量真值掩码与元数据的不足。该数据集包含60张图像,其中30张经过精心伪造的图像(包括简单随机拼接、对象语义拼接和扩散模型修复三种操纵类型)与30张真实图像,每张图像均配有像素级真值掩码和采集元数据。数据集基于fMoW卫星影像库,通过地理配准、尺度对齐和人工辅助的掩码定义等流程创建,确保伪造内容的高真实感。该数据集适用于图像取证算法评估与地理空间深度伪造定位研究,旨在为检测与定位方法提供更具挑战性的标准化测试基准。
fmow-fake-small is a benchmark dataset for satellite image manipulation and deepfake localization, constructed by Oak Ridge National Laboratory. It aims to address the shortcoming that existing remote sensing forgery datasets lack high-quality ground-truth masks and metadata. This dataset contains 60 images in total, including 30 meticulously manipulated fake images covering three manipulation types: simple random splicing, object semantic splicing, and diffusion model inpainting, and 30 real images. Each image is paired with a pixel-level ground-truth mask and acquisition metadata. Built upon the fMoW satellite image repository, this dataset is created through processes such as georegistration, scale alignment, and human-assisted mask definition, ensuring high realism of the forged content. This dataset is applicable to image forensics algorithm evaluation and geospatial deepfake localization research, and aims to provide a more challenging standardized test benchmark for detection and localization methods.




