以人类为中心的伪造数据集

ICCV 2025 HumanSAM: Classifying Human-centric Forgery Videos in Human Spatial, Appearance, and Motion Anomaly HFV 是首个专为**多类别（Multi-class）**以人为中心的视频伪造检测设计的公开基准数据集 。与传统的仅区分“真/假”的二分类数据集不同，HFV 旨在支持更细粒度的伪造理解，将生成视频中的伪造痕迹细分为空间、外观和运动三种特定的异常类型 。

ICCV 2025 HumanSAM: Classifying Human-centric Forgery Videos in Human Spatial, Appearance, and Motion Anomaly. HFV is the first publicly available benchmark dataset specifically designed for multi-class human-centric video forgery detection. Unlike traditional binary classification datasets that only distinguish between "real" and "fake", HFV aims to support more fine-grained forgery analysis, classifying the forgery traces in generated videos into three specific anomaly types: spatial, appearance, and motion anomalies.

搜集汇总

背景与挑战

背景概述

以人类为中心的伪造视频（HFV）数据集是首个公开的多类人类中心视频伪造分类基准，旨在超越传统的真实与伪造二元检测，提供更细粒度的伪造类型理解。该数据集包含四类视频：三类生成内容中的异常类型（空间、外观和运动异常）以及一类真实视频，视频来源涵盖九个先进视频生成模型和Kinetics-400数据集。数据集具有丰富的背景和活动类别，并通过半自动化流程构建，包括异常评分和人工验证。