Replay-Mobile

所有视频均使用移动设备（平板电脑或手机）的前置摄像头拍摄。前置摄像头可生成分辨率为 720 像素（宽）x 1280 像素（高）的彩色视频，并以“.mov”文件格式保存。帧速率约为 25 Hz。真实用户已经执行了真实访问（向设备展示一个人的真实面孔）。通过显示被攻击客户端的照片或视频记录至少 10 秒来执行攻击访问。

在五种不同的照明条件（受控、逆向、直接、横向和漫射）下记录了真实的客户访问。此外，为了产生攻击，每个客户端的高分辨率照片和视频都是在类似于其身份验证会话（lighton、lightoff）的条件下拍摄的。

然后将 1190 个真实访问和攻击视频按以下方式分组：
训练集：包含不同光照条件下的 120 次真实访问和 192 次攻击；
开发集：包含不同光照条件下的160次真实访问和256次攻击；
测试集：包含不同光照条件下的 110 次真实访问和 192 次攻击；
注册集：包含 160 个不同光照条件下的真实访问，**专门**用于研究人脸识别系统的基线性能。 （这个集合再次划分为“训练”、“开发”和“测试”集合。）

All videos were captured using the front-facing camera of a mobile device (tablet or smartphone). The front-facing camera can produce color videos with a resolution of 720 pixels (width) × 1280 pixels (height), and all videos are saved in the ".mov" file format. The frame rate is approximately 25 Hz. Genuine user access was conducted by real users by presenting a live human face to the device. Attack access was executed by displaying photos or video recordings of the attacked client for a minimum of 10 seconds.

Genuine client accesses were recorded under five distinct lighting conditions: controlled, backlight, direct, side, and diffuse. Additionally, to generate attack samples, high-resolution photos and videos of each client were captured under conditions similar to their authentication sessions (lighton, lightoff).

A total of 1190 genuine access and attack videos were partitioned into the following subsets:
- Training set: Contains 120 genuine access samples and 192 attack samples across different lighting conditions;
- Development set: Contains 160 genuine access samples and 256 attack samples across different lighting conditions;
- Test set: Contains 110 genuine access samples and 192 attack samples across different lighting conditions;
- Enrollment set: Contains 160 genuine access samples under different lighting conditions, specifically intended for studying the baseline performance of face recognition systems. This set is further subdivided into "training", "development", and "test" subsets.

搜集汇总

背景与挑战

背景概述

Replay-Mobile是一个用于人脸反欺骗检测的数据集，包含使用移动设备前置摄像头拍摄的视频，分辨率为720x1280，覆盖五种光照条件。数据集划分为训练、开发、测试和注册集，总计包含1190个真实访问和攻击视频，由伊迪亚普研究所在2016年发布。