[总结] 人脸识别领域相关数据集

我就是我 2021-11-17 11:26 615阅读 0赞

> 推荐大家利用这个网站：[Browse state-of-the-art][]，快速了解某个领域包含哪些方向，常用数据集以及SOTA。

### 目录 ###

**1. Introduce**  
**2. Database of Face Detection**  
**3. Database of Face Alignment && Face Landmark Detection**  
**4. Database of Face Recognition && Face Identification && Face Verification**  
**5. Database of Face Attribute Classification**

### 1. Introduce ###

人脸识别领域所包括的方向非常之多，包括：人脸检测、人脸关键点检测、人脸对齐等等；而人们常说的人脸识别严格来说应该是人脸检测（Face Detection）。本文将对人脸识别领域部分相关方向进行总结，以论文小结的形式对各个方向近年的发展趋势进行总结。（GitHub 上有人已经做了不错的总结：[ChanChiChoi/awesome-Face\_Recognition][ChanChiChoi_awesome-Face_Recognition]；想了解各个方向的 SOTA 可以查看 [Browse state-of-the-art][]）

![20190810153642119.png][]

### 2. Database of Face Detection ###

*  **FDDB**
    
    > HomePage : [FDDB: Face Detection Data Set and Benchmark][FDDB_ Face Detection Data Set and Benchmark]  
    > Paper : [FDDB: A Benchmark for Face Detection in Unconstrained Settings][FDDB_ A Benchmark for Face Detection in Unconstrained Settings]  
    > Tasks : FaceFace Detection  
    > 特点：In The Wild，提供椭圆检测框，图像分辨率较小
    
    ![20190826082117917.png][]
    
    该数据集包含 2845 张图像，共计 5171 个人脸，每张图像的标注信息包括：Bounding Box 信息（该数据集除了提供常见的矩形检测框坐标外，还提供椭圆检测框坐标）。此外，该数据集由数据集 [Faces in the Wild][] 挑选组成，还提供了 10 折交叉验证的划分。关于该数据集的评价标准，感兴趣的可以阅读这篇博文：[FDDB人脸检测算法评价标准][FDDB]。
    
    （该数据集较为简单，现在经常作为刷榜的入场资格进行测评，故就不放相应的 SOTA）
 *  **WIDER FACE**
    
    > HomePage : [WIDER FACE: A Face Detection Benchmark][WIDER FACE_ A Face Detection Benchmark]  
    > Paper : [WIDER FACE: A Face Detection Benchmark][WIDER FACE_ A Face Detection Benchmark 1]  
    > Tasks : Face Detection  
    > 特点：In The Wild，数量大，单张图像中存在较多人脸，图像分辨率高
    
    ![20190826081546892.png][]
    
    该数据集包含 32203 张图像，共计 393703 个人脸，每张图像的标注信息包括：Bounding Box 信息，人脸遮挡程度信息，图像光照程度信息等等。该数据集提供训练集、验证集、测试集的划分，数量比例为 50%:10%:40%，且测试集不包含 ground truth，需将测试结果提交至官方主页进行评测。此外，该数据集由数据集 [WIDER Dataset][] 随机挑选组成，还依据检测的难易程度将其划分了 Easy/Medium/Hard 三个难度等级。
    
    State Of The Art (Easy): Accurate Face Detection for High Performance, 2019 ![20190826002210324.png][]State Of The Art (Medium): Accurate Face Detection for High Performance, 2019 ![20190826002245365.png][]State Of The Art (Hard): RetinaFace: Single-stage Dense Face Localisation in the Wild, 2019 ![2019082600233948.png][]

### 3. Database of Face Alignment && Face Landmark Detection ###

Face Alignment 被翻译为面部对齐，即对所检测的人脸图像进行旋正使其符合标准要求，但是往往也包含 Facial Landmark Detection，这两者往往容易混为一谈。以下为两者的定义：

**Face Alignment** : Face alignment is the task of identifying the geometric structure of faces in digital images, and attempting to obtain a canonical alignment of the face based on translation, scale, and rotation.

**Facial Landmark Detection** : Facial landmark detection is the task of detecting key landmarks on the face and tracking them (being robust to rigid and non-rigid facial deformations due to head movements and facial expressions).

*  **AFLW**
    
    > HomePage : [Annotated Facial Landmarks in the Wild (AFLW)][Annotated Facial Landmarks in the Wild _AFLW]  
    > Paper : [Annotated Facial Landmarks in the Wild: A Large-scale, Real-world Database for Facial Landmark Localization][Annotated Facial Landmarks in the Wild_ A Large-scale_ Real-world Database for Facial Landmark Localization]  
    > Tasks : Face Landmark Detection  
    > 特点：In The Wild
    
    AFLW 数据集由 21997 张网络照片组成，共包含 25993 个人脸。该数据集所提供的人脸关键点格式为 21 Facial Landmark，具体细节如下所示
    
    ![20191017155550680.png][]State Of The Art : PFLD: A Practical Facial Landmark Detector, 2019 ![20191017161049262.png][]
 *  **300W**
    
    > HomePage : [300 Faces In-the-Wild Challenge (300-W)][300 Faces In-the-Wild Challenge _300-W]  
    > Paper : [ 300 faces In-the-wild challenge: Database and results][300 faces In-the-wild challenge_ Database and results]  
    > Tasks : Face Landmark Detection, Face Detection  
    > 特点：In The Wild
    
    300-W challenge 所使用的训练数据集实际上由 AFLW，AFW，Helen，IBUG，LFPW，FRGC-V2，XM2VTS等数据集组成，作者将其数据格式进行统一，其人脸关键点格式为 68 Facial Landmark，具体细节如下所示：
    
    ![20191017160756772.png][]State Of The Art : PFLD: A Practical Facial Landmark Detector, 2019 ![20191017160946160.png][]

### 4. Database of Face Recognition && Face Identification && Face Verification ###

虽然 Face Recognition 翻译过来为人脸识别，容易让人混淆，其实用 Face Identification/Verfication 更好一些，即人脸识别/人脸确认。其中，两者的区别如下：

**Face Identification** : Face Identification is 1:many, i.e., you have to find an unknown person in a dataset to answer the question " who is this person". You have to check the biometrics against all others in the database.

**Face Verification** : Face Verification is 1:1, i.e. you have to check if this person is the correct one. In this case you have to know somthing more about the person e.g. name. You only have to check one biometric profile.

**PS:** [ZhaoJ9014/face.evoLVe.PyTorch][ZhaoJ9014_face.evoLVe.PyTorch]

*  **LFW**
    
    > HomePage : [Labeled Faces in the Wild][]  
    > Paper : [Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments][Labeled Faces in the Wild_ A Database for Studying Face Recognition in Unconstrained Environments]  
    > Tasks : Face Identification, Face Verification  
    > 特点：In The Wild
    
    LFW 数据集由 5749 个名人的 13323 张网络照片组成，每张图像都有其唯一的姓名 ID 和序号。其中，有 1680 个人有两个或两个以上不同的照片。
    
    该数据库从中随机选择了 6000 对人脸图像组成了图片对，其中，3000 对属于同一个人的不同人脸照片，3000 对属于不同 ID 的人脸照片。此外，LFW 有部分数据是标注错误的，详细情况请查看其主页的 Errata 部分。
    
    State Of The Art : ArcFace: Additive Angular Margin Loss for Deep Face Recognition, 2018 ![20190922172925874.png][]
 *  **MS-Celeb-1M**
    
    > HomePage : [MS-Celeb-1M][]  
    > Paper : [MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition][MS-Celeb-1M_ A Dataset and Benchmark for Large-Scale Face Recognition]  
    > Tasks : Face Identification, Face Verification  
    > 特点：In The Wild，数量大，官方竞赛
    
    MS-Celeb-1M 由微软所公布，为 Challenge of Recognizing One Million Celebrities in the Real World 的数据集。该数据集包含 10 万个 ID，1000 万张人脸图像，平均每个 ID 有 100 张人脸图像。
 *  **VGGFace**
    
    > HomePage : [VGGFace][]  
    > Paper : [Deep Face Recognition][]  
    > Tasks : Face Identification, Face Verification  
    > 特点：In The Wild，数量大
    
    VGGFace 数据集由牛津大学的 Visual Geometry Group（就是提出 VGG 网络的实验室） 于 2015 年发布，该数据集包括 2622 个 ID，260 万张人脸图像，平均每个 ID 有 1000 张人脸图像。
 *  **VGGFace2**
    
    > HomePage : [VGGFace2][]  
    > Paper : [VGGFace2: A dataset for recognising faces across pose and age][VGGFace2_ A dataset for recognising faces across pose and age]  
    > Tasks : Face Identification, Face Verification  
    > 特点：In The Wild，数量大
    
    VGGFace2 数据集发布于 2018 年，该数据集包括 9131 个 ID，331 万张人脸图像，其中，训练集包含 8631 个 ID，314 万张人脸图像，测试集包含 500 个 ID，17 万张人脸图像。
 *  **MegaFace**
    
    > HomePage : [MegaFace][]  
    > Paper : [The MegaFace Benchmark: 1 Million Faces for Recognition at Scale][The MegaFace Benchmark_ 1 Million Faces for Recognition at Scale]  
    > Tasks : Face Identification, Face Verification  
    > 特点：In The Wild，数量大，官方竞赛
    
    MegaFace 提供两种测评方案：Challenge 1 和 Challenge 2。其中，Challenge 1 要求选手在任意数据集上训练，在 MegaFace 所提供的 Distractors 上测试（该 Distractors 由 690572 个用户的百万张人脸图像组成）；Challenge 2 要求选手在 MegaFace 所提供的训练集上训练（该训练集由 672057 个 ID 的 470 万张人脸图像组成，图像与 ID 之比为 7），在 [FaceScrub][] 和 [FG-Net][] 上进行测试。
    
    State Of The Art : ArcFace: Additive Angular Margin Loss for Deep Face Recognition, 2018 ![2019092217415313.png][]
 *  **IJB-A**
    
    > HomePage : [IARPA Janus Benchmark-A face challenge (IJB-A)][IARPA Janus Benchmark-A face challenge _IJB-A]  
    > Paper : [Pushing the Frontiers of Unconstrained Face Detection and Recognition: IARPA Janus Benchmark A \*][Pushing the Frontiers of Unconstrained Face Detection and Recognition_ IARPA Janus Benchmark A]  
    > Tasks : Face Identification, Face Verification  
    > 特点：官方竞赛
    
    IJB-A 全称为 IARPA Janus Benchmark-A face challenge， 由 [NIST][]（National Institute of Standards and Technology）所提出，该竞赛所提供的数据集包含 500 个对象的 5712 张静态人脸图像和 2085 个人脸视频帧。
    
    State Of The Art (Face Verification): Dual-Agent GANs for Photorealistic and Identity Preserving Profile Face Synthesis, 2017 NIPS ![2019092220113092.png][]State Of The Art (Face Identification) : Pose-Robust Face Recognition via Deep Residual Equivariant Mapping, 2018 ![20190922201603399.png][]
 *  **IJB-B**
    
    > HomePage : [IARPA Janus Benchmark-B face challenge (IJB-B)][IARPA Janus Benchmark-B face challenge _IJB-B]  
    > Tasks : Face Identification, Face Verification  
    > 特点：官方竞赛
    
    IJB-B 由 IJB-A 迭代而来，该竞赛所提供的数据集包含 1845 个对象的 11754 张人脸图像， 55026 个视频帧，7011 个视频，10044 张非人脸图像。
    
    State Of The Art : FacePoseNet: Making a Case for Landmark-Free Face Alignment, 2017 ![20190922201910925.png][]
 *  **IJB-C**
    
    > HomePage : [IARPA Janus Benchmark-C face challenge (IJB-C)][IARPA Janus Benchmark-C face challenge _IJB-C]  
    > Tasks : Face Identification, Face Verification  
    > 特点：官方竞赛
    
    IJB-C 由 IJB-B 迭代而来，该竞赛所提供的数据集包含 3531 个对象的 21294 张人脸图像， 117542 个视频帧，11779 个视频，10040 张非人脸图像。
    
    State Of The Art : Look Across Elapse: Disentangled Representation Learning and Photorealistic Cross-Age Face Synthesis for Age-Invariant Face Recognition, 2018 ![20190922202013314.png][]

### 5. Database of Face Attribute Classification ###

*  **CelebA**
    
    > HomePage : [Large-scale CelebFaces Attributes (CelebA) Dataset][Large-scale CelebFaces Attributes _CelebA_ Dataset]  
    > Paper : [Deep Learning Face Attributes in the Wild][]  
    > Tasks : Face Attribute Recognition, Face Detection, Landmark Localization, and Face Editing & Synthesis.  
    > 特点：In The Wild，数量大，提供多个人脸属性
    
    该数据集包含202599 张图像，共计 10177 个 ID，每张图像的标注信息包括：5 Landmark Locations, 40 Binary Attributes, Identities。此外，该数据集提供原始图像以及经过矫正和裁剪后的图像，提供训练集、验证集、测试集的划分。
    
    State Of The Art : Multi-Task Learning as Multi-Objective Optimization, 2018 ![20190730183228181.png][]
 *  **RAF**
    
    > HomePage : [Real-world Affective Faces Database][]  
    > Paper : [Reliable Crowdsourcing and Deep Locality-Preserving Learning for Expression Recognition in the Wild][]  
    > Tasks : Face Expression Recognition  
    > 特点：In The Wild，数量大
    
    根据相关页面介绍可知，RAF 数据集含有将近 30 K 张图像，并且每张图像标注了 Landmark 信息（准确的 5 landmark 信息，以及由算法检测出的 37 landmark 信息），Bounding Box 信息，种族，年龄范围以及性别信息。此外，该数据集有两个子集，分别包括 7 种 Basic 表情和 12 种 Compound 表情。
    
    State Of The Art : Facial Expression Recognition with Inconsistently Annotated Datasets, 2018 ECCV ![20190802150256153.png][]State Of The Art : Covariance Pooling for Facial Expression Recognition, 2018 CVPR ![20190802150214623.png][]
 *  **SFEW**
    
    > HomePage : [The Third Emotion Recognition in the Wild Challenge (EmotiW 2015)][The Third Emotion Recognition in the Wild Challenge _EmotiW 2015]  
    > Tasks : Face Expression Recognition  
    > 特点：In The Wild
    
    Static Facial Expressions in the Wild Database (SFEW-DB) 为 Acted Facial Expressions In The Wild Databse (AFEW-DB) 的子集。其中，AFEW-DB 由电影中剪辑所得的视频片段组成，作为 Emotion Recognition In The Wild Challenge (EmotiW) 系列情感识别挑战赛所使用的数据集；而 SFEW-DB 由 AFEW-DB 中所抽取的有表情的静态帧组成。
    
    <table> 
     <thead> 
      <tr> 
       <th>Surprise</th> 
       <th>Fear</th> 
       <th>Disgust</th> 
       <th>Happy</th> 
       <th>Sad</th> 
       <th>Angry</th> 
       <th>Neutral</th> 
       <th>Sum</th> 
      </tr> 
     </thead> 
     <tbody> 
      <tr> 
       <td><strong>Train Set</strong></td> 
       <td>96</td> 
       <td>98</td> 
       <td>66</td> 
       <td>198</td> 
       <td>172</td> 
       <td>178</td> 
       <td>150</td> 
      </tr> 
      <tr> 
       <td><strong>Test Set</strong></td> 
       <td>57</td> 
       <td>47</td> 
       <td>23</td> 
       <td>73</td> 
       <td>73</td> 
       <td>77</td> 
       <td>86</td> 
      </tr> 
     </tbody> 
    </table>
    
    State Of The Art : Facial Expression Recognition with Inconsistently Annotated Datasets, 2018 ECCV ![20190802150236709.png][]State Of The Art : Covariance Pooling for Facial Expression Recognition, 2018 CVPR ![20190802150214623.png][]

**参考资料：**

*  [weslynn/AlphaTree-graphic-deep-neural-network][weslynn_AlphaTree-graphic-deep-neural-network]
 *  [ 人脸识别常用数据集大全（12/20更新）][_12_20]
 *  [人脸数据集整理（201804）][201804]
 *  [人脸识别Benchmark：WIDER FACE等相关][Benchmark_WIDER FACE]
 *  [【技术综述】一文道尽“人脸数据集”][Link 1]

![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L0NvZGVfTWFydA_size_16_color_FFFFFF_t_70_pic_center]

[Browse state-of-the-art]: https://paperswithcode.com/sota
[ChanChiChoi_awesome-Face_Recognition]: https://github.com/ChanChiChoi/awesome-Face_Recognition
[20190810153642119.png]: /images/20211116/4359f3cd336f4bcdbd4fbec809db64dc.png
[FDDB_ Face Detection Data Set and Benchmark]: http://vis-www.cs.umass.edu/fddb/
[FDDB_ A Benchmark for Face Detection in Unconstrained Settings]: http://vis-www.cs.umass.edu/fddb/fddb.pdf
[20190826082117917.png]: /images/20211116/71042dee50714f8a9f0c7f20bc19f36c.png
[Faces in the Wild]: http://tamaraberg.com/faceDataset/index.html
[FDDB]: https://yinguobing.com/fddb/
[WIDER FACE_ A Face Detection Benchmark]: http://shuoyang1213.me/WIDERFACE/
[WIDER FACE_ A Face Detection Benchmark 1]: https://arxiv.org/pdf/1511.06523.pdf
[20190826081546892.png]: /images/20211116/25bffdc0f1824c86909a72d1d6c65775.png
[WIDER Dataset]: http://yjxiong.me/event_recog/WIDER/
[20190826002210324.png]: https://img-blog.csdnimg.cn/20190826002210324.png
[20190826002245365.png]: https://img-blog.csdnimg.cn/20190826002245365.png
[2019082600233948.png]: /images/20211116/3bc201b2124a440c9ad4f427e99e3275.png
[Annotated Facial Landmarks in the Wild _AFLW]: https://www.tugraz.at/institute/icg/research/team-bischof/lrs/downloads/aflw/
[Annotated Facial Landmarks in the Wild_ A Large-scale_ Real-world Database for Facial Landmark Localization]: https://files.icg.tugraz.at/seafhttp/files/2a299d28-74f7-4690-9e05-37b130e77f9c/koestinger_befit_11.pdf
[20191017155550680.png]: https://img-blog.csdnimg.cn/20191017155550680.png
[20191017161049262.png]: /images/20211116/42557d525bdc4a6089fe40316d40c1eb.png
[300 Faces In-the-Wild Challenge _300-W]: https://ibug.doc.ic.ac.uk/resources/300-W/
[300 faces In-the-wild challenge_ Database and results]: https://ibug.doc.ic.ac.uk/media/uploads/documents/sagonas_2016_imavis.pdf
[20191017160756772.png]: https://img-blog.csdnimg.cn/20191017160756772.png
[20191017160946160.png]: /images/20211116/35b41bcb666f488fb37023a6a457844e.png
[ZhaoJ9014_face.evoLVe.PyTorch]: https://github.com/ZhaoJ9014/face.evoLVe.PyTorch
[Labeled Faces in the Wild]: http://vis-www.cs.umass.edu/lfw/
[Labeled Faces in the Wild_ A Database for Studying Face Recognition in Unconstrained Environments]: http://vis-www.cs.umass.edu/lfw/lfw.pdf
[20190922172925874.png]: /images/20211116/229547a027cd4df19655af8e53b5fccc.png
[MS-Celeb-1M]: https://www.microsoft.com/en-us/research/project/ms-celeb-1m-challenge-recognizing-one-million-celebrities-real-world/
[MS-Celeb-1M_ A Dataset and Benchmark for Large-Scale Face Recognition]: https://arxiv.org/pdf/1607.08221.pdf
[VGGFace]: http://www.robots.ox.ac.uk/~vgg/data/vgg_face/
[Deep Face Recognition]: http://www.robots.ox.ac.uk/~vgg/publications/2015/Parkhi15/parkhi15.pdf
[VGGFace2]: http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/
[VGGFace2_ A dataset for recognising faces across pose and age]: http://www.robots.ox.ac.uk/~vgg/publications/2018/Cao18/cao18.pdf
[MegaFace]: http://megaface.cs.washington.edu/
[The MegaFace Benchmark_ 1 Million Faces for Recognition at Scale]: https://megaface.cs.washington.edu/KemelmacherMegaFaceCVPR16.pdf
[FaceScrub]: http://vintage.winklerbros.net/facescrub.html
[FG-Net]: https://yanweifu.github.io/FG_NET_data/index.html
[2019092217415313.png]: /images/20211116/a666b382a2374c23a516fd4567da3c2b.png
[IARPA Janus Benchmark-A face challenge _IJB-A]: https://www.nist.gov/itl/iad/image-group/ijb-dataset-request-form
[Pushing the Frontiers of Unconstrained Face Detection and Recognition_ IARPA Janus Benchmark A]: https://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Klare_Pushing_the_Frontiers_2015_CVPR_paper.pdf
[NIST]: https://www.nist.gov
[2019092220113092.png]: https://img-blog.csdnimg.cn/2019092220113092.png
[20190922201603399.png]: /images/20211116/5c763f35d6ee4c80893f69cf23ca9cb8.png
[IARPA Janus Benchmark-B face challenge _IJB-B]: https://www.nist.gov/itl/iad/image-group/ijbb-dataset-request-form
[20190922201910925.png]: /images/20211116/526d717774ac450480beb03cf5759fc6.png
[IARPA Janus Benchmark-C face challenge _IJB-C]: https://www.nist.gov/itl/iad/ig/ijb-c-dataset-request-form
[20190922202013314.png]: /images/20211116/753c72b4addf432a91c724c72eac94b9.png
[Large-scale CelebFaces Attributes _CelebA_ Dataset]: http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
[Deep Learning Face Attributes in the Wild]: https://arxiv.org/pdf/1411.7766
[20190730183228181.png]: /images/20211116/3903f12aaca2472b9e50700ae4812374.png
[Real-world Affective Faces Database]: http://www.whdeng.cn/RAF/model1.html
[Reliable Crowdsourcing and Deep Locality-Preserving Learning for Expression Recognition in the Wild]: http://openaccess.thecvf.com/content_cvpr_2017/papers/Li_Reliable_Crowdsourcing_and_CVPR_2017_paper.pdf
[20190802150256153.png]: https://img-blog.csdnimg.cn/20190802150256153.png
[20190802150214623.png]: /images/20211116/ddf02376a3c24e6b96b20b5da5a13e95.png
[The Third Emotion Recognition in the Wild Challenge _EmotiW 2015]: https://cs.anu.edu.au/few/emotiw2015.html
[20190802150236709.png]: https://img-blog.csdnimg.cn/20190802150236709.png
[weslynn_AlphaTree-graphic-deep-neural-network]: https://github.com/weslynn/AlphaTree-graphic-deep-neural-network/blob/master/database/face.md
[_12_20]: https://zhuanlan.zhihu.com/p/31378836
[201804]: https://blog.csdn.net/Martin20150405/article/details/79788501
[Benchmark_WIDER FACE]: https://blog.csdn.net/qq_24118527/article/details/84673364
[Link 1]: https://zhuanlan.zhihu.com/p/48347016
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L0NvZGVfTWFydA_size_16_color_FFFFFF_t_70_pic_center]: /images/20211116/fea914d3f13d4fd194f0b4a30eb7bee6.png