dejian-lc/HFV

人类中心伪造视频（HFV）数据集是首个为多类人类中心视频伪造分类任务构建的公开基准数据集。它超越了传统的二元（真实与伪造）检测，提供了对伪造类型的细粒度理解，这对于现实应用中的可靠性和可解释性至关重要。数据集包含4个类别：三种不同的异常类型和一个真实视频类。视频来源包括由九种最先进的视频生成模型生成的合成视频和来自Kinetics-400数据集的真实视频。数据集具有多样性，包含约20种背景和80种人类活动类别，所有视频分辨率均保证≥224×224。

The Human-centric Forgery Video (HFV) Dataset is the first public benchmark constructed for the task of multi-class human-centric video forgery classification. Moving beyond traditional binary (real vs. fake) detection, HFV aims to provide a fine-grained understanding of forgery types, which is critical for reliability and interpretability in real-world applications. The dataset contains 4 classes: three distinct anomaly types observed in generated content, and one class for real videos. The dataset is comprehensive, including synthetic videos generated by nine state-of-the-art video generation models as well as real videos sourced from the Kinetics-400 (K400) dataset. It includes approximately 20 types of backgrounds and 80 categories of human activities, with all video resolutions guaranteed to be ≥224×224.