【数据集介绍】The Idiap Research Institute REPLAY-Mobile Database

【时间】2018.11.14

概述

本文是对数据集REPLAY-Mobile Database中的readme.md文件的中英对照翻译，是对这个数据库的内容与使用方法的介绍。事先声明，本人无法提供该数据库，万分抱歉，如果需要，请到官网申请并下载。官网链接为：https://www.idiap.ch/dataset/replay-mobile/download-proc

The Replay-Mobile Database

==========================

一、简介

The Replay-Mobile Database for face spoofing consists of 1190 video clips of photo and video attack attempts to 40 clients, under different lighting conditions. These videos have been recorded with current devices from the market: an iPad Mini2 tablet and a LG-G4 smartphone. This database has been produced at the Idiap Research Institute (Switzerland) within the framework of collaboration with Galician Research and Development Center in Advanced Telecommunications. Gradiant (Spain).

【用于人脸反欺诈的the Replay-Mobile Database包含尝试使用照片和视频进行攻击的1190个视频片段，它们实在不同的灯光条件下进行拍摄的。这些视频都是用市场上现有的设备录制的：一个iPad微型2平板电脑和一个LG-G4智能手机。该数据库是在与加利西亚先进电信研究和开发中心合作的框架内在idiap研究所（瑞士）制作的。】

Note:

The database would have consisted of 1200 videos. For client009 (test subset), however, we were not able to collect real-access videos (the person came for only the first recording session, not the second session). Therefore, for this client, we have only videos for the enrollment and the attacks.

【注意：该数据库将包含1200个视频。然而，对于client009（测试子集），我们无法收集实际访问视频（该人仅来自第一次录制会话，而不是第二次会话）。因此，对于此客户，我们只有视频用于注册和攻击。】

二、欺诈攻击描述

Spoofing Attacks Description

----------------------------

This 2D face spoofing attack database consists of 1,190 video clips of photo and video attack attempts of 40 clients, under various lighting conditions.

【这个2D面部欺骗攻击数据库包含1,190个视频剪辑，包括40个客户在各种照明条件下的照片和视频攻击尝试。】

The data is split into 4 sub-groups:

* Training data ("train"), to be used for training your anti-spoof classifier;

* Development data ("devel"), to be used for threshold estimation;

* Test data ("test"), with which to report error figures;

* Enrollment data ("enroll"), that can be used to verify spoofing sensitivity on face detection algorithms.

【数据分为4个子组：

*训练数据（“训练”），用于训练您的反欺骗分类器;

*用于阈值估计的开发数据（“devel”）;

*测试数据（“测试”），用于报告错误数据;

*注册数据（“注册”），可用于验证面部检测算法的欺骗灵敏度。】

The data-sets 'train', 'devel', and 'test' are disjoint. Clients that appear in one of these data sets (train, devel or test) do not appear in the other two sets.

【数据集'train'，'devel'和'test'是不相交的。出现在其中一个数据集（train，devel或test）中的客户端不会出现在其他两个数据集中。】

三、数据集描述

Database Description

--------------------

All videos are captured using the front-camera of the mobile device (tablet or mobile). The front-camera produces colour videos with a high definition resolution of 720 pixels (width) by 1280 pixels (height) which are saved in ".mov" format. The frame rate is about 25 Hz. Real accesses have been performed using the face of the genuine user. Attacks attempts have been performed by displaying a photo or a video recording of the same client for at least 10 seconds.

【使用移动设备（平板电脑或移动设备）的前置摄像头捕获所有视频。前置摄像头产生的彩色视频具有720像素（宽度）×1280像素（高度）的高清晰度分辨率，以“.mov”格式保存。帧速率约为25 Hz。我们使用真实用户的面部执行了真实访问。通过显示同一客户端的照片或视频记录至少10秒来执行攻击尝试。】

3.1光照条件

Real client accesses have been recorded under five different lighting conditions:

* **controlled** : The office light turned on, blinds down, background homogeneous

* **adverse** : The office light turned off, blinds halfway up, background homogeneous

* **direct** : The user captured the video in front of a window with direct sunlight, with more complex background.

* **lateral** : The user captured the video perpendicular to the window with lateral sunlight, with more complex background.

* **diffuse** : The user captured the video in an open hall with diffuse illumination, with more complex background.

【在五种不同的照明条件下记录了真实客户访问：

* **控制**：办公室灯亮，百叶窗关闭，背景均匀

* **不利**：办公室灯关闭，百叶窗半开，背景均匀

* **直接**：用户在阳光直射的窗前拍摄视频，背景更复杂。

* **横向**：用户在侧面阳光下垂直于窗户拍摄视频，背景更复杂。

* **漫反射**：用户在开放式大厅中拍摄视频，漫射照明，背景更复杂。】

【5种灯光方式加上2种拍摄设备，所以每个客户有10个real视频，如下】

To produce the attacks, high-resolution photos and videos from each client have been used under similar conditions as in their authentication sessions.

* **lighton** : The user sitting and the capturer device was holder on a tripod, the office light was turned on, blinds up, background is homogeneous

* **lightoff** : The user sitting and the capturer device was holder on a tripod, the office light was turned off, blinds up, background is homogeneous

【为了产生攻击，来自每个客户端的高分辨率照片和视频在与其身份验证会话类似的条件下使用。

* ** 灯亮 **：用户坐着，捕获器设备放在三脚架上，办公室灯打开，百叶窗打开，背景均匀

* **灯灭**：用户坐着，捕捉器装置放在三脚架上，办公室灯关闭，百叶窗打开，背景均匀】

3.2 攻击数据的获取

For photos attacks a Nikon Coolpix P520 camera has been used. The images have been captured with 18 Mpixel resolution. Video attacks have been captured using the back-camera of a LG-G4 smartphone, which records 1080p FHD video clips through it's 16 Mpixel camera.

【对于照片攻击，使用了尼康Coolpix P520相机。图像以18 Mpixel分辨率拍摄。而使用LG-G4智能手机的后置摄像头捕获视频攻击，该智能手机通过其1600万像素摄像头记录1080p FHD视频片段。】

Two kinds of attacks have been performed:

* **mattescreen** : All the attacks have been displayed on Philips 227ELH screen (resolution 1920x1080 pixels). This matte screen avoids reflections. The videos have been recorded using devices supported on a stand. Two kinds of attacks have been performed using mattescreens;

* **photo**: a still photo of the attacked identity is displayed on the screen.

* **video**: a video showing the attacked identity is replayed on the screen.

* **print**: Al the attacks have been printed on Konica Minolta ineo+ 224e color laser printer. The videos have been recorded in two modes:

* **fixed** : Using devices supported on a stand.

* **hand** : Using devices held by the spoofer.

In total, 16 attack videos have been recorded for each client, 8 for each of the attacking modes described above.

【已经进行了两种攻击：

* ** 磨砂显示屏 **：所有攻击都已在飞利浦227ELH屏幕上显示（分辨率为1920x1080像素）。这种磨砂屏幕避免了反射。视频已使用支架上支撑的设备进行录制。使用mattescreens进行了两种攻击;

* **照片**：屏幕上显示受攻击身份的静态照片。

* **视频**：在屏幕上重播显示受攻击身份的视频。

* **打印**：这些攻击已经由柯尼卡美能达ineo + 224e彩色激光打印机打印。

视频以两种模式录制：

* **固定**：使用支架上支撑的设备。

* **手**：使用欺骗者持有的设备。

总共为每个客户端记录了16个攻击视频，对于上述每种攻击模式记录了8个。】

3.3.攻击展示（display）类别

* 4 x mobile attacks using a mattescreen displaying:

* 1 x mobile photo/lighton

* 1 x mobile photo/lightoff

* 1 x mobile video/lighton

* 1 x mobile video/lightoff

* 4 x tablet attacks using a mattescreen displaying:

* 1 x tablet photo/lighton

* 1 x tablet photo/lightoff

* 1 x tablet video/lighton

* 1 x tablet video/lightoff

* 2 x Print attacks captured by smartphone with fixed support. The print image occupied the entire available printing surface on A4 paper for the following samples:

* 1 x high-resolution print of photo/lighton

* 1 x high-resolution print of photo/lightoff

* 2 x Print attacks captured by tablet with fixed support. The print image occupied the entire available printing surface on A4 paper for the following samples:

* 1 x high-resolution print of photo/lighton

* 1 x high-resolution print of photo/lightoff

* 2 x Print attacks captured by hand-held smartphone. The print image occupied the entire available printing surface on A4 paper for the following samples:

* 1 x high-resolution print of photo/lighton

* 1 x high-resolution print of photo/lightoff

* 2 x Print attacks captured by hand-held tablet. The print image occupied the entire available printing surface on A4 paper for the following samples:

* 1 x high-resolution print of photo/lighton

* 1 x high-resolution print of photo/lightoff

【每个客户的16种攻击视频举例如下：】

【*使用磨砂屏显示4种移动攻击：

* 1 x移动照片/ lighton

* 1 x移动照片/lightoff

* 1 x移动视频/ lighton

* 1 x移动视频/lightoff

*使用磨砂屏显示的4种平板电脑攻击：

* 1 x平板电脑照片/ lighton

* 1 x平板电脑照片/lightoff

* 1 x平板电脑视频/ lighton

* 1 x平板电脑视频/lightoff

*具有固定支持的智能手机捕获的2 种打印攻击。打印图像占据A4纸上的整个可用打印表面，用于以下样本：

* 1张高分辨率照片/lighton

* 1张高分辨率照片/lightoff

*具有固定支持的平板电脑捕获的2 种打印攻击。打印图像占据A4纸上的整个可用打印表面，用于以下样本：

* 1张高分辨率照片/lighton

* 1张高分辨率照片/lightoff

* 2个手持智能手机拍摄的打印攻击。打印图像占据A4纸上的整个可用打印表面，用于以下样本：

* 1张高分辨率照片/lighton

* 1张高分辨率照片/lightoff

*手持平板电脑拍摄的2 x打印攻击。打印图像占据A4纸上的整个可用打印表面，用于以下样本：

* 1张高分辨率照片/lighton

* 1张高分辨率照片/lightoff】

The following images illustrate the set-up for capturing videos of attacks using a matte screen.

![Alt](images/mattescreen_attack_1.jpg)

![Alt](images/mattescreen_attack_2.jpg)

![Alt](images/mattescreen_attack_3.jpg)

The images below show how print-attack videos have been captured.

![Alt](images/print_attack_1.jpg)

![Alt](images/print_attack_2.jpg)

【以下图像说明了使用磨砂屏幕捕获攻击视频的设置。

！[ALT]（图像/ mattescreen_attack_1.jpg）

！[ALT]（图像/ mattescreen_attack_2.jpg）

！[ALT]（图像/ mattescreen_attack_3.jpg）

下图显示了如何捕获打印攻击视频。

！[ALT]（图像/ print_attack_1.jpg）

！[ALT]（图像/ print_attack_2.jpg）】

3.4 数据集的划分

The 1200 real-accesses and attacks videos were then divided in the following way:

* **Training set**: contains 120 real-accesses and 192 attacks under different lighting conditions;

* **Development set**: contains 160 real-accesses and 256 attacks under different lighting conditions;

* **Test set**: contains 110 real-accesses and 192 attacks under different lighting conditions;

* **Enrollment set**: contains 160 real-accesses under various lighting conditions, to be used **exclusively** for studying the baseline performance of face recognition systems.

【按以下方式划分1200个实际访问和攻击视频：

* **训练集**：在不同的光照条件下包含120次实际访问和192次攻击;

* **开发集**：在不同光照条件下包含160次实际访问和256次攻击;

* **测试集**：在不同光照条件下包含110次实际访问和192次攻击;

* **注册集**：在各种照明条件下包含160个实际访问，仅用于**用于研究人脸识别系统的基线性能。】

四、人脸定位

Face Locations

--------------

We also provide face location (bounding-boxes) automatically annotated by a cascade of classifier based on a variant of Local Binary Patterns (LBP).

The bob package [bob.ip.facedetect](https://github.com/bioidiap/bob.ip.facedetect) has been used to generate the face-locations.

For each video, face-location is computed for each frame, and face-locations for all frames of a video are stored in a single file in a 4-column format (x,y,width, height). For each video, two versions of the face-location file are provided, one in ASCII text format (extension: .face), and the other in hdf5 format (extension: .hdf5).

The face-location files can be found in the folder `database/faceloc`.

【我们还提供基于局部二进制模式（LBP）的变体由级联分类器自动注释的面部位置（边界框）。

bob包[bob.ip.facedetect]（https://github.com/bioidiap/bob.ip.facedetect）已用于生成面部位置。

对于每个视频，计算每个帧的面部位置，并将视频的所有帧的面部位置存储在4列的单个文件中格式（x，y，宽度，高度）。对于每个视频，提供两个版本的面部位置文件，一个是ASCII文本格式（扩展名：.face），和另一种是hdf5格式（扩展名：.hdf5）。

面部位置文件可以在`database / faceloc`文件夹中找到。

】

五、合法生物识别交易协议

Protocol for Licit Biometric Transactions

-----------------------------------------

It is possible to measure the performance of baseline face recognition systems on the 2D Face spoofing database and to evaluate how well the attacks confound such systems. Here we describe how to use the available data at the enrollment set to create a background model, client models and how to perform scoring using the available data.

1. Universal Background Model (UBM): To generate the UBM, select the training-set client videos from the enrollment videos. There should be 4 per client (12 clients), that is, a total of 48 videos, each with 300 frames to create the model;

2. Client models: To generate client models, use the enrollment data for clients in the development and test groups. Again, there should be 4 videos per client (one for each light condition and one for each device). At the end of the enrollment procedure, the development set must have 1 model for each of the 16 clients available in that set, and similarly, for the test set, 1 model must be generated for each of the 12 clients available;

3. For a simple baseline verification, generate scores **exhaustively** for all videos from the development and test **real-accesses** respectively, but **without** intermixing across development and test sets. The scores generated against matched client videos and models (within the subset, i.e. development or test) should be considered true client accesses, while all others should be considered impostors;

4、To obtain a single number to quantify the performance, do the following: exclusively using the scores from the development-set, tune your face recognition system to achieve the lowest possible EER for the development-set; then use the score-threshold corresponding to this EER, to find the HTER on the test-set scores.

【可以在2D Face欺骗数据库上测量基线面部识别系统的性能，并评估攻击对这些系统的混淆程度。在这里，我们将介绍如何使用注册集中的可用数据来创建背景模型，客户端模型以及如何使用可用数据执行评分。

1.通用背景模型（UBM）：要生成UBM，请从注册视频中选择训练集客户端视频。每个客户端应该有4个（12个客户端），即总共48个视频，每个视频有300帧来创建模型;

2.客户端模型：要生成客户端模型，请使用开发和测试组中的客户端的注册数据。同样，每个客户端应该有4个视频（每个光照条件一个，每个设备一个）。在注册过程结束时，开发集必须为该组中可用的16个客户端中的每个客户端提供1个模型，同样，对于测试集，必须为可用的12个客户端中的每个客户端生成1个模型;

3.对于简单的基线验证，分别为开发和测试**实际访问**中的所有视频生成** ** **，但**不会在开发和测试集之间进行**混合。针对匹配的客户端视频和模型（在子集内，即开发或测试）生成的分数应被视为真正的客户端访问，而所有其他视频和模型应被视为冒名顶替者;

4.要获得单个数字来量化性能，请执行以下操作：仅使用开发集中的分数，调整人脸识别系统以实现开发集的尽可能低的EER;然后使用与此EER相对应的分数阈值，在测试集分数上找到HTER。】

六、欺骗攻击的协议

Protocols for Spoofing Attacks

------------------------------

Attack protocols are used to evaluate the (binary classification) performance of counter-measures to spoof attacks. Replay-mobile supports 7 different protocols (see table below) according to the type of device used to generate the attack: print, mattescreen. Furthermore, subsetting can be achieved on top of these 7 groups by classifying attacks performed by hand-held devices or using devices on a fixed support. This classification scheme makes-up a total of 21 protocols that can be used for studying the performance of counter-measures to 2D face spoofing attacks. The table below details the number of video clips in each protocol.

Distribution of videos per attack-protocol in the Replay-Mobile 2D Face spoofing database.

【攻击协议用于评估针对欺骗攻击的反措施的（二进制分类）性能。 Replay-mobile根据用于生成攻击的设备类型支持7种不同的协议（见下表）：print，mattescreen。此外，通过对手持设备执行的攻击或固定支持上的设备进行分类，可以在这7个组之上实现子集化。该分类方案构成总共21种协议，可用于研究对抗2D面部欺骗攻击的对策的性能。下表详细列出了每种协议中的视频片段数量。

在Replay-Mobile 2D Face欺骗数据库中按攻击协议分发视频如下。

======================= ===== ==== ===== ===== ==== ====== ===== ==== =====

Mobile Tablet All Devices

---------------- ---------------- ----------------

Protocol train dev test train dev test train dev test

paper-allsupport 48 48 64 48 48 64 96 128 96

- paper-hand 24 24 32 24 24 32 48 64 48

- paper-fixed 24 24 32 24 24 32 48 64 48

mattescreen-alltype 48 48 64 48 48 64 96 128 96

- mattescreen-video 24 24 32 24 24 32 48 64 48

- mattescreen-photo 24 24 32 24 24 32 48 64 48

Grandtest 96 96 128 96 96 128 192 256 192

======================== ===== ===== ===== ===== ==== ====== ===== ==== =====