Seq-Deepfake

Seq-DeepFake是第一个用于顺序DeepFake操纵检测的大规模数据集。它由85k顺序处理的人脸图像组成，每个图像都带有地面真相序列注释。数据集包括长度从0到5的高多样性操作序列，并且基于两种不同的面部操作方法生成:
顺序面部组件操作
面部图像的35,166数量
28种操作序列
顺序面部属性操作
面部图像的49,920数量
26种操作序列的类型

Seq-DeepFake is the first large-scale dataset tailored for sequential DeepFake manipulation detection. It consists of 85,000 sequentially processed facial images, each paired with ground-truth sequence annotations. The dataset includes highly diverse manipulation sequences with lengths ranging from 0 to 5, and is generated using two distinct facial manipulation methodologies:
- Sequential Facial Component Manipulation: 35,166 facial images, 28 types of manipulation sequences
- Sequential Facial Attribute Manipulation: 49,920 facial images, 26 types of manipulation sequences

搜集汇总

背景与挑战

背景概述

Seq-Deepfake是一个用于顺序DeepFake操纵检测的大规模数据集，包含85k顺序处理的人脸图像，涵盖两种面部操作方法和多种操作序列长度。该数据集由南洋理工大学实验室于2022年发布。