Identification Rate vs. Distractors Size


Algorithm Date Submitted Set 1 Set 2 Set 3 Data Set Size
THU CV-AI Lab 12/12/2017 77.977% 77.995% 77.968% Large
EI Networks 8/10/2018 75.974% 75.974% 75.974% Large
Google - FaceNet v8 10/23/2015 74.594% 74.585% 74.558% Large
QINIU ATLAB - FaceX V1 (iBUG cleaned data) 7/23/2018 72.438% 72.438% 72.438% Large
SIATMMLAB TencentVision 12/1/2016 71.247% 71.283% 71.256% Large
SRC-Beijing-FR(Samsung Research Institute China-Beijing) 8/15/2018 70.0% 70.0% 70.0% Large
BingMMLab V1(iBUG cleaned data) 4/10/2018 69.109% 69.109% 69.109% Large
BingMMLab-v1 (non-cleaned data) 4/10/2018 69.108% 69.108% 69.108% Large
ATLAB-FACEX (QINIU CLOUD) 6/23/2018 68.59% 68.59% 68.59% Large
TencentAILab_FaceCNN_v1 9/21/2017 67.873% 67.855% 67.918% Large
iBUG_DeepInsight 2/8/2018 67.476% 67.512% 67.575% Large
iBug (Reported by Author) 04/28/2017 66.94% Small
FaceTag V1 12/18/2017 66.086% 66.185% 66.149% Large
Yang Sun 06/05/2017 65.491% 65.599% 65.563% Large
ULUFace 5/7/2018 64.6878% 64.6878% 64.6878% Large
DeepSense V2 1/22/2017 63.632% 63.632% 63.632% Large
YouTu Lab (Tencent Best-Image) 04/08/2017 61.611% 61.693% 61.6711% Large
Vocord - deepVo V3 04/27/2017 61.395% 61.386% 61.431% Large
SenseTime PureFace(clean) 6/13/2018 59.003% 59.003% 59.003% Large
CVTE V2 1/27/2018 57.163% 57.191% 57.182% Small
Fiberhome AI Research Lab V2(Nanjing) 8/28/2018 57.0462% 57.0462% 57.0462% Large
SIAT_MMLAB 3/29/2016 55.304% 53.41% 53.049% Small
Intellivision 2/11/2018 54.836% 54.944% 54.908% Large
NTechLAB - facenx_large 10/20/2015 52.716% 52.733% 52.743% Large
Video++ 1/5/2018 52.03% 52.048% 52.003% Large
Faceter Lab 12/18/2017 49.729% 49.792% 49.738% Large
DeepSense - Large 07/31/2016 49.368% 49.396% 49.549% Large
Progressor 09/13/2017 49.26% 49.26% 49.332% Large
SphereFace - Small 12/1/2016 47.555% 47.582% 47.591% Small
TUPUTECH 12/22/2017 47.248% 47.384% 47.194% Large
Shanghai Tech 08/13/2016 45.209% 45.182% 45.209% Large
Vocord-deepVo1.2 12/1/2016 44.966% 45.047% 45.029% Large
DeepSense - Small 07/31/2016 43.54% 43.576% 43.522% Small
CVTE 08/30/2017 42.349% 42.431% 42.431% Large
3DiVi Company - tdvm V2 04/15/2017 40.256% 40.166% 40.22% Large
ForceInfo 04/07/2017 39.859% 39.949% 39.895% Large
Vocord - DeepVo1 08/03/2016 39.489% 39.471% 39.534% Large
Beijing Faceall Co. - FaceAll V2 04/28/2017 39.038% 38.975% 39.02% Small
XT-tech V2 08/30/2017 36.855% 36.945% 36.918% Large
NTechLAB - facenx_small 10/20/2015 29.168% 29.15% 29.168% Small
Beijing Faceall Co. - FaceAll_1600 10/19/2015 26.155% 26.155% 26.137% Large
Fudan University - FUDAN-CS_SDS 1/29/2017 25.568% 25.559% 25.623% Small
Beijing Faceall Co. - FaceAll_Norm_1600 10/19/2015 25.02% 25.045% 24.973% Large
GRCCV 12/1/2016 24.783% 24.802% 24.783% Small
3DiVi Company - tdvm6 10/27/2015 15.78% 15.825% 15.77% Small
Barebones_FR - cnn 10/21/2015 7.136% 10.709% 10.736% Small

Method Details

Algorithm Details
Orion Star Technology (clean)

We have trained three deep networks (ResNet-101, ResNet-152, ResNet-200) with joint softmax and triplet loss on MS-Celeb-1M (95K identities, 5.1M images), and the triplet part is trained by batch online hard negative mining with subspace learning. The features of all networks are concatenated to produce the final feature, whose dimension is set to be 256x3. For data processing, we use original large images and follow our own system by detection and alignment. Particularly, in evaluation, we have cleaned the FaceScrub and MegaFace with the code released by iBUG_DeepInsight.

Orion Star Technology (no clean)

Compared to our another submission named “Orion Star Technology (clean)”, the major difference is that no any data cleaning is adopted in evaluation.

SuningUS_AILab

This is a model ensembled by three different models using ResNet CNN and improved ResNet network, learned by a combined loss. A filtered MS-Celeb-1M and CASIA-Webface is used as the dataset.

ULSee - Face Team

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

Large-Margin Softmax Loss for Convolutional Neural Networks https://arxiv.org/abs/1612.02295

A Discriminative Deep Feature Learning Approach for Face Recognition

https://ydwen.github.io/papers/WenECCV16.pdf

NormFace: L2 Hypersphere Embedding for Face Verification

https://arxiv.org/abs/1704.06369

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

SIATMMLAB TencentVision

Adopt the ensemble of very deep CNNs, learned by joint supervision (softmax loss, improved center loss, etc). Training data is a combination of public datasets (CAISA, VGG, CACD2000, etc) and private datasets. The total number of images is more than 2 million.

DeepSense - Small

Adopt a network of very deep ResNet CNNs, learned by combined supervision(identification loss(softmax loss), verification loss, triplet loss).

GRCCV

The algorithm consists of three parts: FCN - based fast face detection algorithm, pre-training ResNet CNN on classification task, weight tuning. Training set contains 273000 photos. Hardware: 8 x Tesla k80.

SphereFace - Small

SphereFace uses a novel approach to learn face features that are discriminative on a hypersphere manifold. The training data set we use in SphereFace is the publicly available CASIA-WebFace dataset which contains 490k images of nearly 10,500 individuals.

EM-DATA

arcface, https://github.com/deepinsight/insightface

StartDT-AI

we only use a single model trained on a ResNet-28 network joined with cosine loss and triplet loss on MS-Celeb-1M(74K identities, 4M images), refer to

DeepVisage: Making face recognition simple yet with powerful generalization skills

https://arxiv.org/abs/1703.08388

One-shot Face Recognition by Promoting Underrepresented Classes

https://arxiv.org/abs/1707.05574v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

BingMMLab-v1 (non-cleaned data)

Compares to our submissions named “BingMMLab V1(cleaned data)”, the only difference is that no data cleaning is adopted in this evaluation.

BingMMLab V1(iBUG cleaned data)

We used knowledge graph to collect identities and then crawled Bing search engine to get high quality images, we filtered noises in 14M training data by clustering with weak face models, and then trained a DensetNet-69 (k=48) network with A-Softmax loss variants. In evaluation, we used the cleaned test set released by iBUG_DeepInsight.

DenseNet

https://arxiv.org/pdf/1608.06993v2.pdf

A-Softmax and its variant:

https://arxiv.org/pdf/1704.08063.pdf

https://github.com/wy1iu/LargeMargin_Softmax_Loss/issues/13

MTDP_ITC(Clean)

Angular Softmax Loss with Channel-wise Attention

Kankan AI Lab

We have trained our model on ResNet-152 with Additive Angular Margin Loss on combined dataset with MS-Celeb-1M and VggFace2, and cleaned the FaceScrub and MegaFace with the lists released by iBUG_DeepInsight.

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

https://github.com/deepinsight/insightface

VGGFace2: A dataset for recognising faces across pose and age

https://arxiv.org/abs/1710.08092

MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition

https://www.microsoft.com/en-us/research/wp-content/uploads/2016/08/MSCeleb-1M-a.pdf

TUPUTECH V1 (iBUG cleaned data)

We have trained ResNet models with a combined loss on MS-Celeb-1M. In evaluation, we have cleaned the FaceScrub and MegaFace with the code released by iBUG_DeepInsight.

[iBUG_DeepInsight code](https://github.com/deepinsight/insightface)

TUPUTECH v2

Compares to our submissions named “TUPUTECH v1 (clean)”, the only difference is data cleaning by TUPU is adopted in evaluation.

Uface

trained network with arcface loss. tried some different methods in preprocessing data.In evaluation, used the cleaned test set released by iBUG_DeepInsight.

ULUFace

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

MTCNN:Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

FaceNet: A Unified Embedding for Face Recognition and Clustering

https://arxiv.org/abs/1503.03832

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

ULUFace

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

MTCNN:Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

FaceNet: A Unified Embedding for Face Recognition and Clustering

https://arxiv.org/abs/1503.03832

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

Jian24 Vision

we only use a single model trained on a ResNet-101 network joined with cosine loss and triplet loss on MS-Celeb-1M(74K identities, 4M images), refer to

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

A Discriminative Deep Feature Learning Approach for Face Recognition

https://ydwen.github.io/papers/WenECCV16.pdf

sophon

using insightface with loss modification.

MSU_Intsys

Custom version of ArcFace.

FeelingTech

FeelingFace, which trained with cleaned MS-Celeb-1M dataset

KANKAN AI Lab

We have trained our model on ResNet-152 with Additive Angular Margin Loss on combined dataset with MS-Celeb-1M and VggFace2, and cleaned the FaceScrub and MegaFace with the lists released by iBUG_DeepInsight.

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

https://github.com/deepinsight/insightface

VGGFace2: A dataset for recognising faces across pose and age

https://arxiv.org/abs/1710.08092

MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition

https://www.microsoft.com/en-us/research/wp-content/uploads/2016/08/MSCeleb-1M-a.pdf

Sogou

We collected and filtered millions pictures from sogou pic search engine, and also used pictures date cleaned by DeepInsight. We trained multi model by different loss functions such as combined margin loss, margin loss with focal loss and so on. In some models we divided faces into different patches so we can get features represent local feature such as mouth and even teeth. We merge the features get from different models together as the final result.

SenseTime PureFace(clean)

We trained the deep residual attention network(attention-56) with A-softmax to learn the face feature.

The training dataset is constructed by the novel dataset building techinique, which is critical for us to improve the performance of the model. The results are the cleaned test set performance released by iBUG_DeepInsight.

Wang F, Jiang M, Qian C, et al. Residual attention network for image classification[J]. arXiv preprint arXiv:1704.06904, 2017.

Deng J, Guo J, Zafeiriou S. ArcFace: Additive Angular Margin Loss for Deep Face Recognition[J]. arXiv preprint arXiv:1801.07698, 2018.

Liu W, Wen Y, Yu Z, et al. Sphereface: Deep hypersphere embedding for face recognition[C]//The IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017, 1.

EM-DATA

arcface, https://github.com/deepinsight/insightface

ATLAB-FACEX (QINIU CLOUD)

A single deep Resnet model (the 'r100' configuration from insightface) trained with our own angular margin loss (an improved variant of A-Softmax, not published yet).

The training set consists of nearly 6.2M images (about 96K identities). For MegaFace evaluation, we adopted the clean list released by iBUG_DeepInsight.

insightface: https://github.com/deepinsight/insightface

A-Softmax: https://arxiv.org/pdf/1704.08063.pdf

Qsdream

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

PingAn AI Lab (Nanjing)

This is a single model trained by a deep ResNet network on MS-Celeb-1M,learned by a cosine loss, refer to ArcFace: Additive Angular Margin Loss for Deep Face Recognition https://arxiv.org/abs/1801.07698 SphereFace: Deep Hypersphere Embedding for Face Recognition https://arxiv.org/abs/1704.08063 We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

Fiberhome AI Research Lab(Nanjing)

mtcnn: Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

https://github.com/pangyupo/mxnet_mtcnn_face_detection

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

https://github.com/deepinsight/insightface

QINIU ATLAB - FaceX V1 (iBUG cleaned data)

This feature model is an ensemble of 3 deep Resnet models (the 'r100' and 'r152' configuration from insightface[1] ) trained with our own angular margin loss (an improved variant of A-Softmax[2] , not published yet).

The training set consists of 7 Million images (about 188K identities). For evaluation, we adopted the clean list released by iBUG_DeepInsight[1] .

Reference:

[1] insightface: https://github.com/deepinsight/insightface

[2] A-Softmax: https://arxiv.org/pdf/1704.08063.pdf

Visual Computing-Alibaba-V1(clean)

A single model (improved Resnet-152) is trained by the supervision of combined loss functions (A-Softmax loss, center loss, triplet loss et al) on MS-Celeb-1M (84 k identities, 5.2 M images). In evaluation, we use the cleaned FaceScrub and MegaFace released by iBUG_DeepInsight.

Beijing Faceall Co. & BUPT(iBug cleaned)

We have trained on the Faceall-msra celebrities dataset with over 4.4 million photos. We use 6 models to ensemble the training result and use cosine distance as the distance metrics. As for loss function, we adopt A-softmax and Additive Angular margin loss during training. We evaluate our result on the iBUG-cleaned version of megaface and facescrub list.

Method reference:

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

EI Networks

We build a training database of 120,000 identities and 12 million images with combination of public and private databases. We remove the overlap with Facescrub and Fgnet database from our training set. Three deep residual networks are trained (one resnet-150 like and two resnet-100 like) on 112x96 input image with multiple large margin loss functions. Each network is further finetuned using triplet loss. Output feature of three networks is concatenated and trained with metric learning for dimension reduction to a vector of size 512.

Uniview Technology

use the fusion model of resnet and googlenet

SRC-Beijing-FR(Samsung Research Institute China-Beijing)

An improved loss of sphereface with a large-scale training dataset.

4paradigm

We have trained resnet101 model with large additive margin softmax loss on merged MS-Celeb-1M and Asian-Celeb and fine-tune the model with batchhard triplet loss . In evaluation, we cleaned the FaceScrub and MegaFace using noisy face images released by[1]

[1]Deng J, Guo J, Zafeiriou S. ArcFace: Additive Angular Margin Loss for Deep Face Recognition[J]. 2018.

4paradigm

We have trained resnet101 model with combine large margin softmax loss on merged MS-Celeb-1M and Asian-Celeb and fine-tune the model with batchhard triplet loss . In evaluation, we cleaned the FaceScrub and MegaFace using the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

Fiberhome AI Research Lab V2(Nanjing)

Compares to our submissions named “Fiberhome AI Research Lab ”,the differences are the following:

1.we changed the face alignment method.

2.we added private datasets to train

3.we adopted a RestNet-50 network joined with cosine loss and Additive Angular Margin Loss

CyberLink

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://kpzhang93.github.io/MTCNN_face_detection_alignment/index.html

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v4

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063v4

We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

CyberLink_mobile

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://kpzhang93.github.io/MTCNN_face_detection_alignment/index.html

MobileFaceNets: Efficient CNNs for Accurate Real-Time Face Verification on Mobile Devices

https://arxiv.org/abs/1804.07573v4

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v4

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063v4

We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

cyberlink_resnet-v2

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://kpzhang93.github.io/MTCNN_face_detection_alignment/index.html

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063v4

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v4

We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

Sogou AIGROUP - SFace

We collected and filtered millions pictures from sogou pic search engine and mining one hundred thousand hard negative sample, all data are cleaned by DeepInsight. We trained multi model by different loss functions such as combined margin loss, margin loss with focal loss and so on. In some models we divided faces into different patches so we can get features represent local feature, especially the tooth similarity model. We merge the features get from different models together as the final result.

ICARE_FACE_V1

We have trained our model based on a deep convolutional neural network(ResNet101) with Additive Margin Softmax.We have semi-automatically cleaned the training dataset MSCeleb-1M.Particularly,in evaluation,we cleaned the FaceScrub and MegaFace with the lists released by iBUG_DeepInsight.

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599

Insightface clean list

https://github.com/deepinsight/insightface

Rank-1 Identification Performance

Rank-10 Identification Performance





JointBayes Verification Graph

- -   uses large training set

Set 1

JointBayes Verification Graph

Set 1

JointBayes Verification Graph






JointBayes Verification Graph

- -   uses large training set

Set 2

JointBayes Verification Graph

Set 2

JointBayes Verification Graph






JointBayes Verification Graph

- -   uses large training set

Set 3

JointBayes Verification Graph

Set 3

JointBayes Verification Graph

Identification Rank vs. Rank


Algorithm Date Submitted Set 1 Set 2 Set 3 Data Set Size
THU CV-AI Lab 12/12/2017 77.977% 77.995% 77.968% Large
EI Networks 8/10/2018 75.974% 75.974% 75.974% Large
Google - FaceNet v8 10/23/2015 74.594% 74.585% 74.558% Large
QINIU ATLAB - FaceX V1 (iBUG cleaned data) 7/23/2018 72.438% 72.438% 72.438% Large
SIATMMLAB TencentVision 12/1/2016 71.247% 71.283% 71.256% Large
SRC-Beijing-FR(Samsung Research Institute China-Beijing) 8/15/2018 70.0% 70.0% 70.0% Large
BingMMLab V1(iBUG cleaned data) 4/10/2018 69.109% 69.109% 69.109% Large
BingMMLab-v1 (non-cleaned data) 4/10/2018 69.108% 69.108% 69.108% Large
ATLAB-FACEX (QINIU CLOUD) 6/23/2018 68.59% 68.59% 68.59% Large
TencentAILab_FaceCNN_v1 9/21/2017 67.873% 67.855% 67.918% Large
iBUG_DeepInsight 2/8/2018 67.476% 67.512% 67.575% Large
iBug (Reported by Author) 04/28/2017 66.94% Small
FaceTag V1 12/18/2017 66.086% 66.185% 66.149% Large
Yang Sun 06/05/2017 65.491% 65.599% 65.563% Large
ULUFace 5/7/2018 64.6878% 64.6878% 64.6878% Large
DeepSense V2 1/22/2017 63.632% 63.632% 63.632% Large
YouTu Lab (Tencent Best-Image) 04/08/2017 61.611% 61.693% 61.6711% Large
Vocord - deepVo V3 04/27/2017 61.395% 61.386% 61.431% Large
SenseTime PureFace(clean) 6/13/2018 59.003% 59.003% 59.003% Large
CVTE V2 1/27/2018 57.163% 57.191% 57.182% Small
Fiberhome AI Research Lab V2(Nanjing) 8/28/2018 57.0462% 57.0462% 57.0462% Large
SIAT_MMLAB 3/29/2016 55.304% 53.41% 53.049% Small
Intellivision 2/11/2018 54.836% 54.944% 54.908% Large
NTechLAB - facenx_large 10/20/2015 52.716% 52.733% 52.743% Large
Video++ 1/5/2018 52.03% 52.048% 52.003% Large
Faceter Lab 12/18/2017 49.729% 49.792% 49.738% Large
DeepSense - Large 07/31/2016 49.368% 49.396% 49.549% Large
Progressor 09/13/2017 49.26% 49.26% 49.332% Large
SphereFace - Small 12/1/2016 47.555% 47.582% 47.591% Small
TUPUTECH 12/22/2017 47.248% 47.384% 47.194% Large
Shanghai Tech 08/13/2016 45.209% 45.182% 45.209% Large
Vocord-deepVo1.2 12/1/2016 44.966% 45.047% 45.029% Large
DeepSense - Small 07/31/2016 43.54% 43.576% 43.522% Small
CVTE 08/30/2017 42.349% 42.431% 42.431% Large
3DiVi Company - tdvm V2 04/15/2017 40.256% 40.166% 40.22% Large
ForceInfo 04/07/2017 39.859% 39.949% 39.895% Large
Vocord - DeepVo1 08/03/2016 39.489% 39.471% 39.534% Large
Beijing Faceall Co. - FaceAll V2 04/28/2017 39.038% 38.975% 39.02% Small
XT-tech V2 08/30/2017 36.855% 36.945% 36.918% Large
NTechLAB - facenx_small 10/20/2015 29.168% 29.15% 29.168% Small
Beijing Faceall Co. - FaceAll_1600 10/19/2015 26.155% 26.155% 26.137% Large
Fudan University - FUDAN-CS_SDS 1/29/2017 25.568% 25.559% 25.623% Small
Beijing Faceall Co. - FaceAll_Norm_1600 10/19/2015 25.02% 25.045% 24.973% Large
GRCCV 12/1/2016 24.783% 24.802% 24.783% Small
3DiVi Company - tdvm6 10/27/2015 15.78% 15.825% 15.77% Small
Barebones_FR - cnn 10/21/2015 7.136% 10.709% 10.736% Small

Method Details

Algorithm Details
Orion Star Technology (clean)

We have trained three deep networks (ResNet-101, ResNet-152, ResNet-200) with joint softmax and triplet loss on MS-Celeb-1M (95K identities, 5.1M images), and the triplet part is trained by batch online hard negative mining with subspace learning. The features of all networks are concatenated to produce the final feature, whose dimension is set to be 256x3. For data processing, we use original large images and follow our own system by detection and alignment. Particularly, in evaluation, we have cleaned the FaceScrub and MegaFace with the code released by iBUG_DeepInsight.

Orion Star Technology (no clean)

Compared to our another submission named “Orion Star Technology (clean)”, the major difference is that no any data cleaning is adopted in evaluation.

SuningUS_AILab

This is a model ensembled by three different models using ResNet CNN and improved ResNet network, learned by a combined loss. A filtered MS-Celeb-1M and CASIA-Webface is used as the dataset.

ULSee - Face Team

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

Large-Margin Softmax Loss for Convolutional Neural Networks https://arxiv.org/abs/1612.02295

A Discriminative Deep Feature Learning Approach for Face Recognition

https://ydwen.github.io/papers/WenECCV16.pdf

NormFace: L2 Hypersphere Embedding for Face Verification

https://arxiv.org/abs/1704.06369

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

SIATMMLAB TencentVision

Adopt the ensemble of very deep CNNs, learned by joint supervision (softmax loss, improved center loss, etc). Training data is a combination of public datasets (CAISA, VGG, CACD2000, etc) and private datasets. The total number of images is more than 2 million.

DeepSense - Small

Adopt a network of very deep ResNet CNNs, learned by combined supervision(identification loss(softmax loss), verification loss, triplet loss).

GRCCV

The algorithm consists of three parts: FCN - based fast face detection algorithm, pre-training ResNet CNN on classification task, weight tuning. Training set contains 273000 photos. Hardware: 8 x Tesla k80.

SphereFace - Small

SphereFace uses a novel approach to learn face features that are discriminative on a hypersphere manifold. The training data set we use in SphereFace is the publicly available CASIA-WebFace dataset which contains 490k images of nearly 10,500 individuals.

EM-DATA

arcface, https://github.com/deepinsight/insightface

StartDT-AI

we only use a single model trained on a ResNet-28 network joined with cosine loss and triplet loss on MS-Celeb-1M(74K identities, 4M images), refer to

DeepVisage: Making face recognition simple yet with powerful generalization skills

https://arxiv.org/abs/1703.08388

One-shot Face Recognition by Promoting Underrepresented Classes

https://arxiv.org/abs/1707.05574v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

BingMMLab-v1 (non-cleaned data)

Compares to our submissions named “BingMMLab V1(cleaned data)”, the only difference is that no data cleaning is adopted in this evaluation.

BingMMLab V1(iBUG cleaned data)

We used knowledge graph to collect identities and then crawled Bing search engine to get high quality images, we filtered noises in 14M training data by clustering with weak face models, and then trained a DensetNet-69 (k=48) network with A-Softmax loss variants. In evaluation, we used the cleaned test set released by iBUG_DeepInsight.

DenseNet

https://arxiv.org/pdf/1608.06993v2.pdf

A-Softmax and its variant:

https://arxiv.org/pdf/1704.08063.pdf

https://github.com/wy1iu/LargeMargin_Softmax_Loss/issues/13

MTDP_ITC(Clean)

Angular Softmax Loss with Channel-wise Attention

Kankan AI Lab

We have trained our model on ResNet-152 with Additive Angular Margin Loss on combined dataset with MS-Celeb-1M and VggFace2, and cleaned the FaceScrub and MegaFace with the lists released by iBUG_DeepInsight.

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

https://github.com/deepinsight/insightface

VGGFace2: A dataset for recognising faces across pose and age

https://arxiv.org/abs/1710.08092

MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition

https://www.microsoft.com/en-us/research/wp-content/uploads/2016/08/MSCeleb-1M-a.pdf

TUPUTECH V1 (iBUG cleaned data)

We have trained ResNet models with a combined loss on MS-Celeb-1M. In evaluation, we have cleaned the FaceScrub and MegaFace with the code released by iBUG_DeepInsight.

[iBUG_DeepInsight code](https://github.com/deepinsight/insightface)

TUPUTECH v2

Compares to our submissions named “TUPUTECH v1 (clean)”, the only difference is data cleaning by TUPU is adopted in evaluation.

Uface

trained network with arcface loss. tried some different methods in preprocessing data.In evaluation, used the cleaned test set released by iBUG_DeepInsight.

ULUFace

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

MTCNN:Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

FaceNet: A Unified Embedding for Face Recognition and Clustering

https://arxiv.org/abs/1503.03832

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

ULUFace

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

MTCNN:Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

FaceNet: A Unified Embedding for Face Recognition and Clustering

https://arxiv.org/abs/1503.03832

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

Jian24 Vision

we only use a single model trained on a ResNet-101 network joined with cosine loss and triplet loss on MS-Celeb-1M(74K identities, 4M images), refer to

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

A Discriminative Deep Feature Learning Approach for Face Recognition

https://ydwen.github.io/papers/WenECCV16.pdf

sophon

using insightface with loss modification.

MSU_Intsys

Custom version of ArcFace.

FeelingTech

FeelingFace, which trained with cleaned MS-Celeb-1M dataset

KANKAN AI Lab

We have trained our model on ResNet-152 with Additive Angular Margin Loss on combined dataset with MS-Celeb-1M and VggFace2, and cleaned the FaceScrub and MegaFace with the lists released by iBUG_DeepInsight.

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

https://github.com/deepinsight/insightface

VGGFace2: A dataset for recognising faces across pose and age

https://arxiv.org/abs/1710.08092

MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition

https://www.microsoft.com/en-us/research/wp-content/uploads/2016/08/MSCeleb-1M-a.pdf

Sogou

We collected and filtered millions pictures from sogou pic search engine, and also used pictures date cleaned by DeepInsight. We trained multi model by different loss functions such as combined margin loss, margin loss with focal loss and so on. In some models we divided faces into different patches so we can get features represent local feature such as mouth and even teeth. We merge the features get from different models together as the final result.

SenseTime PureFace(clean)

We trained the deep residual attention network(attention-56) with A-softmax to learn the face feature.

The training dataset is constructed by the novel dataset building techinique, which is critical for us to improve the performance of the model. The results are the cleaned test set performance released by iBUG_DeepInsight.

Wang F, Jiang M, Qian C, et al. Residual attention network for image classification[J]. arXiv preprint arXiv:1704.06904, 2017.

Deng J, Guo J, Zafeiriou S. ArcFace: Additive Angular Margin Loss for Deep Face Recognition[J]. arXiv preprint arXiv:1801.07698, 2018.

Liu W, Wen Y, Yu Z, et al. Sphereface: Deep hypersphere embedding for face recognition[C]//The IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017, 1.

EM-DATA

arcface, https://github.com/deepinsight/insightface

ATLAB-FACEX (QINIU CLOUD)

A single deep Resnet model (the 'r100' configuration from insightface) trained with our own angular margin loss (an improved variant of A-Softmax, not published yet).

The training set consists of nearly 6.2M images (about 96K identities). For MegaFace evaluation, we adopted the clean list released by iBUG_DeepInsight.

insightface: https://github.com/deepinsight/insightface

A-Softmax: https://arxiv.org/pdf/1704.08063.pdf

Qsdream

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

PingAn AI Lab (Nanjing)

This is a single model trained by a deep ResNet network on MS-Celeb-1M,learned by a cosine loss, refer to ArcFace: Additive Angular Margin Loss for Deep Face Recognition https://arxiv.org/abs/1801.07698 SphereFace: Deep Hypersphere Embedding for Face Recognition https://arxiv.org/abs/1704.08063 We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

Fiberhome AI Research Lab(Nanjing)

mtcnn: Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

https://github.com/pangyupo/mxnet_mtcnn_face_detection

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

https://github.com/deepinsight/insightface

QINIU ATLAB - FaceX V1 (iBUG cleaned data)

This feature model is an ensemble of 3 deep Resnet models (the 'r100' and 'r152' configuration from insightface[1] ) trained with our own angular margin loss (an improved variant of A-Softmax[2] , not published yet).

The training set consists of 7 Million images (about 188K identities). For evaluation, we adopted the clean list released by iBUG_DeepInsight[1] .

Reference:

[1] insightface: https://github.com/deepinsight/insightface

[2] A-Softmax: https://arxiv.org/pdf/1704.08063.pdf

Visual Computing-Alibaba-V1(clean)

A single model (improved Resnet-152) is trained by the supervision of combined loss functions (A-Softmax loss, center loss, triplet loss et al) on MS-Celeb-1M (84 k identities, 5.2 M images). In evaluation, we use the cleaned FaceScrub and MegaFace released by iBUG_DeepInsight.

Beijing Faceall Co. & BUPT(iBug cleaned)

We have trained on the Faceall-msra celebrities dataset with over 4.4 million photos. We use 6 models to ensemble the training result and use cosine distance as the distance metrics. As for loss function, we adopt A-softmax and Additive Angular margin loss during training. We evaluate our result on the iBUG-cleaned version of megaface and facescrub list.

Method reference:

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

EI Networks

We build a training database of 120,000 identities and 12 million images with combination of public and private databases. We remove the overlap with Facescrub and Fgnet database from our training set. Three deep residual networks are trained (one resnet-150 like and two resnet-100 like) on 112x96 input image with multiple large margin loss functions. Each network is further finetuned using triplet loss. Output feature of three networks is concatenated and trained with metric learning for dimension reduction to a vector of size 512.

Uniview Technology

use the fusion model of resnet and googlenet

SRC-Beijing-FR(Samsung Research Institute China-Beijing)

An improved loss of sphereface with a large-scale training dataset.

4paradigm

We have trained resnet101 model with large additive margin softmax loss on merged MS-Celeb-1M and Asian-Celeb and fine-tune the model with batchhard triplet loss . In evaluation, we cleaned the FaceScrub and MegaFace using noisy face images released by[1]

[1]Deng J, Guo J, Zafeiriou S. ArcFace: Additive Angular Margin Loss for Deep Face Recognition[J]. 2018.

4paradigm

We have trained resnet101 model with combine large margin softmax loss on merged MS-Celeb-1M and Asian-Celeb and fine-tune the model with batchhard triplet loss . In evaluation, we cleaned the FaceScrub and MegaFace using the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

Fiberhome AI Research Lab V2(Nanjing)

Compares to our submissions named “Fiberhome AI Research Lab ”,the differences are the following:

1.we changed the face alignment method.

2.we added private datasets to train

3.we adopted a RestNet-50 network joined with cosine loss and Additive Angular Margin Loss

CyberLink

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://kpzhang93.github.io/MTCNN_face_detection_alignment/index.html

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v4

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063v4

We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

CyberLink_mobile

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://kpzhang93.github.io/MTCNN_face_detection_alignment/index.html

MobileFaceNets: Efficient CNNs for Accurate Real-Time Face Verification on Mobile Devices

https://arxiv.org/abs/1804.07573v4

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v4

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063v4

We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

cyberlink_resnet-v2

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://kpzhang93.github.io/MTCNN_face_detection_alignment/index.html

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063v4

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v4

We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

Sogou AIGROUP - SFace

We collected and filtered millions pictures from sogou pic search engine and mining one hundred thousand hard negative sample, all data are cleaned by DeepInsight. We trained multi model by different loss functions such as combined margin loss, margin loss with focal loss and so on. In some models we divided faces into different patches so we can get features represent local feature, especially the tooth similarity model. We merge the features get from different models together as the final result.

ICARE_FACE_V1

We have trained our model based on a deep convolutional neural network(ResNet101) with Additive Margin Softmax.We have semi-automatically cleaned the training dataset MSCeleb-1M.Particularly,in evaluation,we cleaned the FaceScrub and MegaFace with the lists released by iBUG_DeepInsight.

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599

Insightface clean list

https://github.com/deepinsight/insightface

Identification Performance with 1 Million Distractors

Identification Performance with 10K Distractors





JointBayes Verification Graph

- -   uses large training set

Set 1

JointBayes Verification Graph

Set 1

JointBayes Verification Graph






JointBayes Verification Graph

- -   uses large training set

Set 2

JointBayes Verification Graph

Set 2

JointBayes Verification Graph






JointBayes Verification Graph

- -   uses large training set

Set 3

JointBayes Verification Graph

Set 3

JointBayes Verification Graph

Verification


Algorithm Date Submitted Set 1 Set 2 Set 3 Data Set Size
Google - FaceNet v8 10/23/2015 75.55% 75.55% 75.55% Large
EI Networks 8/10/2018 70.119% 70.119% 70.119% Large
QINIU ATLAB - FaceX V1 (iBUG cleaned data) 7/23/2018 69.957% 69.957% 69.957% Large
SRC-Beijing-FR(Samsung Research Institute China-Beijing) 8/15/2018 69.0% 69.0% 69.0% Large
THU CV-AI Lab 12/12/2017 68.261% 68.513% 68.188% Large
SIATMMLAB TencentVision 12/1/2016 67.954% 67.954% 67.954% Large
FaceTag V1 12/18/2017 65.789% 65.789% 65.789% Large
TencentAILab_FaceCNN_v1 9/21/2017 64.886% 64.724% 64.724% Large
ULUFace 5/7/2018 64.6878% 64.6878% 64.6878% Large
iBUG_DeepInsight 2/8/2018 64.291% 64.291% 64.291% Large
Yang Sun 06/05/2017 63.623% 64.688% 65.247% Large
ATLAB-FACEX (QINIU CLOUD) 6/23/2018 61.29% 61.29% 61.29% Large
BingMMLab V1(iBUG cleaned data) 4/10/2018 59.78% 59.78% 59.78% Large
BingMMLab-v1 (non-cleaned data) 4/10/2018 59.78% 59.78% 59.78% Large
DeepSense V2 1/22/2017 56.767% 56.767% 56.767% Large
SenseTime PureFace(clean) 6/13/2018 54.835% 54.835% 54.835% Large
Fiberhome AI Research Lab V2(Nanjing) 8/28/2018 54.529% 54.529% 54.529% Large
YouTu Lab (Tencent Best-Image) 04/08/2017 53.681% 53.681% 53.681% Large
Vocord - deepVo V3 04/27/2017 53.573% 53.573% 54.637% Large
CVTE V2 1/27/2018 53.338% 53.338% 53.338% Small
SIAT_MMLAB 3/29/2016 50.144% 51.155% 51.155% Small
TUPUTECH 12/22/2017 45.976% 45.976% 45.976% Large
NTechLAB - facenx_large 10/20/2015 45.381% 45.507% 44.37% Large
iBug (Reported by Author) 04/28/2017 44.947% Small
Intellivision 2/11/2018 43.793% 44.894% 43.793% Large
Vocord-deepVo1.2 12/1/2016 43.252% 43.252% 43.252% Large
SphereFace - Small 12/1/2016 40.094% 40.094% 40.094% Small
Video++ 1/5/2018 35.709% 34.645% 34.645% Large
Vocord - DeepVo1 08/03/2016 35.709% 35.709% 35.709% Large
3DiVi Company - tdvm V2 04/15/2017 33.075% 33.075% 33.075% Large
CVTE 08/30/2017 32.281% 32.064% 32.064% Large
Faceter Lab 12/18/2017 30.982% 30.982% 30.982% Large
Progressor 09/13/2017 29.61% 29.61% 29.61% Large
DeepSense - Small 07/31/2016 29.61% 29.61% 29.61% Small
DeepSense - Large 07/31/2016 29.177% 29.177% 29.177% Large
Shanghai Tech 08/13/2016 27.463% 26.416% 26.416% Large
Beijing Faceall Co. - FaceAll V2 04/28/2017 26.54% 26.54% 26.54% Small
GRCCV 12/1/2016 22.068% 22.068% 22.068% Small
Beijing Faceall Co. - FaceAll_Norm_1600 10/19/2015 18.243% 18.243% 18.243% Large
NTechLAB - facenx_small 10/20/2015 16.961% 16.961% 16.961% Small
Fudan University - FUDAN-CS_SDS 1/29/2017 16.528% 16.528% 16.528% Small
ForceInfo 04/07/2017 16.059% 16.059% 16.059% Large
Beijing Faceall Co. - FaceAll_1600 10/19/2015 15.843% 15.843% 15.843% Large
XT-tech V2 08/30/2017 13.587% 13.587% 13.587% Large
3DiVi Company - tdvm6 10/27/2015 8.336% 8.336% 8.336% Small
Barebones_FR - cnn 10/21/2015 4.764% 5.792% 5.792% Small

Method Details

Algorithm Details
Orion Star Technology (clean)

We have trained three deep networks (ResNet-101, ResNet-152, ResNet-200) with joint softmax and triplet loss on MS-Celeb-1M (95K identities, 5.1M images), and the triplet part is trained by batch online hard negative mining with subspace learning. The features of all networks are concatenated to produce the final feature, whose dimension is set to be 256x3. For data processing, we use original large images and follow our own system by detection and alignment. Particularly, in evaluation, we have cleaned the FaceScrub and MegaFace with the code released by iBUG_DeepInsight.

Orion Star Technology (no clean)

Compared to our another submission named “Orion Star Technology (clean)”, the major difference is that no any data cleaning is adopted in evaluation.

SuningUS_AILab

This is a model ensembled by three different models using ResNet CNN and improved ResNet network, learned by a combined loss. A filtered MS-Celeb-1M and CASIA-Webface is used as the dataset.

ULSee - Face Team

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

Large-Margin Softmax Loss for Convolutional Neural Networks https://arxiv.org/abs/1612.02295

A Discriminative Deep Feature Learning Approach for Face Recognition

https://ydwen.github.io/papers/WenECCV16.pdf

NormFace: L2 Hypersphere Embedding for Face Verification

https://arxiv.org/abs/1704.06369

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

SIATMMLAB TencentVision

Adopt the ensemble of very deep CNNs, learned by joint supervision (softmax loss, improved center loss, etc). Training data is a combination of public datasets (CAISA, VGG, CACD2000, etc) and private datasets. The total number of images is more than 2 million.

DeepSense - Small

Adopt a network of very deep ResNet CNNs, learned by combined supervision(identification loss(softmax loss), verification loss, triplet loss).

GRCCV

The algorithm consists of three parts: FCN - based fast face detection algorithm, pre-training ResNet CNN on classification task, weight tuning. Training set contains 273000 photos. Hardware: 8 x Tesla k80.

SphereFace - Small

SphereFace uses a novel approach to learn face features that are discriminative on a hypersphere manifold. The training data set we use in SphereFace is the publicly available CASIA-WebFace dataset which contains 490k images of nearly 10,500 individuals.

EM-DATA

arcface, https://github.com/deepinsight/insightface

StartDT-AI

we only use a single model trained on a ResNet-28 network joined with cosine loss and triplet loss on MS-Celeb-1M(74K identities, 4M images), refer to

DeepVisage: Making face recognition simple yet with powerful generalization skills

https://arxiv.org/abs/1703.08388

One-shot Face Recognition by Promoting Underrepresented Classes

https://arxiv.org/abs/1707.05574v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

BingMMLab-v1 (non-cleaned data)

Compares to our submissions named “BingMMLab V1(cleaned data)”, the only difference is that no data cleaning is adopted in this evaluation.

BingMMLab V1(iBUG cleaned data)

We used knowledge graph to collect identities and then crawled Bing search engine to get high quality images, we filtered noises in 14M training data by clustering with weak face models, and then trained a DensetNet-69 (k=48) network with A-Softmax loss variants. In evaluation, we used the cleaned test set released by iBUG_DeepInsight.

DenseNet

https://arxiv.org/pdf/1608.06993v2.pdf

A-Softmax and its variant:

https://arxiv.org/pdf/1704.08063.pdf

https://github.com/wy1iu/LargeMargin_Softmax_Loss/issues/13

MTDP_ITC(Clean)

Angular Softmax Loss with Channel-wise Attention

Kankan AI Lab

We have trained our model on ResNet-152 with Additive Angular Margin Loss on combined dataset with MS-Celeb-1M and VggFace2, and cleaned the FaceScrub and MegaFace with the lists released by iBUG_DeepInsight.

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

https://github.com/deepinsight/insightface

VGGFace2: A dataset for recognising faces across pose and age

https://arxiv.org/abs/1710.08092

MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition

https://www.microsoft.com/en-us/research/wp-content/uploads/2016/08/MSCeleb-1M-a.pdf

TUPUTECH V1 (iBUG cleaned data)

We have trained ResNet models with a combined loss on MS-Celeb-1M. In evaluation, we have cleaned the FaceScrub and MegaFace with the code released by iBUG_DeepInsight.

[iBUG_DeepInsight code](https://github.com/deepinsight/insightface)

TUPUTECH v2

Compares to our submissions named “TUPUTECH v1 (clean)”, the only difference is data cleaning by TUPU is adopted in evaluation.

Uface

trained network with arcface loss. tried some different methods in preprocessing data.In evaluation, used the cleaned test set released by iBUG_DeepInsight.

ULUFace

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

MTCNN:Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

FaceNet: A Unified Embedding for Face Recognition and Clustering

https://arxiv.org/abs/1503.03832

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

ULUFace

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

MTCNN:Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

FaceNet: A Unified Embedding for Face Recognition and Clustering

https://arxiv.org/abs/1503.03832

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

Jian24 Vision

we only use a single model trained on a ResNet-101 network joined with cosine loss and triplet loss on MS-Celeb-1M(74K identities, 4M images), refer to

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

A Discriminative Deep Feature Learning Approach for Face Recognition

https://ydwen.github.io/papers/WenECCV16.pdf

sophon

using insightface with loss modification.

MSU_Intsys

Custom version of ArcFace.

FeelingTech

FeelingFace, which trained with cleaned MS-Celeb-1M dataset

KANKAN AI Lab

We have trained our model on ResNet-152 with Additive Angular Margin Loss on combined dataset with MS-Celeb-1M and VggFace2, and cleaned the FaceScrub and MegaFace with the lists released by iBUG_DeepInsight.

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

https://github.com/deepinsight/insightface

VGGFace2: A dataset for recognising faces across pose and age

https://arxiv.org/abs/1710.08092

MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition

https://www.microsoft.com/en-us/research/wp-content/uploads/2016/08/MSCeleb-1M-a.pdf

Sogou

We collected and filtered millions pictures from sogou pic search engine, and also used pictures date cleaned by DeepInsight. We trained multi model by different loss functions such as combined margin loss, margin loss with focal loss and so on. In some models we divided faces into different patches so we can get features represent local feature such as mouth and even teeth. We merge the features get from different models together as the final result.

SenseTime PureFace(clean)

We trained the deep residual attention network(attention-56) with A-softmax to learn the face feature.

The training dataset is constructed by the novel dataset building techinique, which is critical for us to improve the performance of the model. The results are the cleaned test set performance released by iBUG_DeepInsight.

Wang F, Jiang M, Qian C, et al. Residual attention network for image classification[J]. arXiv preprint arXiv:1704.06904, 2017.

Deng J, Guo J, Zafeiriou S. ArcFace: Additive Angular Margin Loss for Deep Face Recognition[J]. arXiv preprint arXiv:1801.07698, 2018.

Liu W, Wen Y, Yu Z, et al. Sphereface: Deep hypersphere embedding for face recognition[C]//The IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017, 1.

EM-DATA

arcface, https://github.com/deepinsight/insightface

ATLAB-FACEX (QINIU CLOUD)

A single deep Resnet model (the 'r100' configuration from insightface) trained with our own angular margin loss (an improved variant of A-Softmax, not published yet).

The training set consists of nearly 6.2M images (about 96K identities). For MegaFace evaluation, we adopted the clean list released by iBUG_DeepInsight.

insightface: https://github.com/deepinsight/insightface

A-Softmax: https://arxiv.org/pdf/1704.08063.pdf

Qsdream

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

PingAn AI Lab (Nanjing)

This is a single model trained by a deep ResNet network on MS-Celeb-1M,learned by a cosine loss, refer to ArcFace: Additive Angular Margin Loss for Deep Face Recognition https://arxiv.org/abs/1801.07698 SphereFace: Deep Hypersphere Embedding for Face Recognition https://arxiv.org/abs/1704.08063 We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

Fiberhome AI Research Lab(Nanjing)

mtcnn: Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://arxiv.org/abs/1604.02878

https://github.com/pangyupo/mxnet_mtcnn_face_detection

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698

https://github.com/deepinsight/insightface

QINIU ATLAB - FaceX V1 (iBUG cleaned data)

This feature model is an ensemble of 3 deep Resnet models (the 'r100' and 'r152' configuration from insightface[1] ) trained with our own angular margin loss (an improved variant of A-Softmax[2] , not published yet).

The training set consists of 7 Million images (about 188K identities). For evaluation, we adopted the clean list released by iBUG_DeepInsight[1] .

Reference:

[1] insightface: https://github.com/deepinsight/insightface

[2] A-Softmax: https://arxiv.org/pdf/1704.08063.pdf

Visual Computing-Alibaba-V1(clean)

A single model (improved Resnet-152) is trained by the supervision of combined loss functions (A-Softmax loss, center loss, triplet loss et al) on MS-Celeb-1M (84 k identities, 5.2 M images). In evaluation, we use the cleaned FaceScrub and MegaFace released by iBUG_DeepInsight.

Beijing Faceall Co. & BUPT(iBug cleaned)

We have trained on the Faceall-msra celebrities dataset with over 4.4 million photos. We use 6 models to ensemble the training result and use cosine distance as the distance metrics. As for loss function, we adopt A-softmax and Additive Angular margin loss during training. We evaluate our result on the iBUG-cleaned version of megaface and facescrub list.

Method reference:

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v2

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

EI Networks

We build a training database of 120,000 identities and 12 million images with combination of public and private databases. We remove the overlap with Facescrub and Fgnet database from our training set. Three deep residual networks are trained (one resnet-150 like and two resnet-100 like) on 112x96 input image with multiple large margin loss functions. Each network is further finetuned using triplet loss. Output feature of three networks is concatenated and trained with metric learning for dimension reduction to a vector of size 512.

Uniview Technology

use the fusion model of resnet and googlenet

SRC-Beijing-FR(Samsung Research Institute China-Beijing)

An improved loss of sphereface with a large-scale training dataset.

4paradigm

We have trained resnet101 model with large additive margin softmax loss on merged MS-Celeb-1M and Asian-Celeb and fine-tune the model with batchhard triplet loss . In evaluation, we cleaned the FaceScrub and MegaFace using noisy face images released by[1]

[1]Deng J, Guo J, Zafeiriou S. ArcFace: Additive Angular Margin Loss for Deep Face Recognition[J]. 2018.

4paradigm

We have trained resnet101 model with combine large margin softmax loss on merged MS-Celeb-1M and Asian-Celeb and fine-tune the model with batchhard triplet loss . In evaluation, we cleaned the FaceScrub and MegaFace using the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

Fiberhome AI Research Lab V2(Nanjing)

Compares to our submissions named “Fiberhome AI Research Lab ”,the differences are the following:

1.we changed the face alignment method.

2.we added private datasets to train

3.we adopted a RestNet-50 network joined with cosine loss and Additive Angular Margin Loss

CyberLink

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://kpzhang93.github.io/MTCNN_face_detection_alignment/index.html

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v4

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063v4

We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

CyberLink_mobile

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://kpzhang93.github.io/MTCNN_face_detection_alignment/index.html

MobileFaceNets: Efficient CNNs for Accurate Real-Time Face Verification on Mobile Devices

https://arxiv.org/abs/1804.07573v4

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v4

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063v4

We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

cyberlink_resnet-v2

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks

https://kpzhang93.github.io/MTCNN_face_detection_alignment/index.html

Deep Residual Learning for Image Recognition

https://arxiv.org/abs/1512.03385v1

ArcFace: Additive Angular Margin Loss for Deep Face Recognition

https://arxiv.org/abs/1801.07698v1

SphereFace: Deep Hypersphere Embedding for Face Recognition

https://arxiv.org/abs/1704.08063v4

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599v4

We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface

Sogou AIGROUP - SFace

We collected and filtered millions pictures from sogou pic search engine and mining one hundred thousand hard negative sample, all data are cleaned by DeepInsight. We trained multi model by different loss functions such as combined margin loss, margin loss with focal loss and so on. In some models we divided faces into different patches so we can get features represent local feature, especially the tooth similarity model. We merge the features get from different models together as the final result.

ICARE_FACE_V1

We have trained our model based on a deep convolutional neural network(ResNet101) with Additive Margin Softmax.We have semi-automatically cleaned the training dataset MSCeleb-1M.Particularly,in evaluation,we cleaned the FaceScrub and MegaFace with the lists released by iBUG_DeepInsight.

Additive Margin Softmax for Face Verification

https://arxiv.org/abs/1801.05599

Insightface clean list

https://github.com/deepinsight/insightface

Verification Performance with 1 Million Distractors

Verifification Performance with 10K Distractors





JointBayes Verification Graph

- -   uses large training set

Set 1

JointBayes Verification Graph

Set 1

JointBayes Verification Graph






JointBayes Verification Graph

- -   uses large training set

Set 2

JointBayes Verification Graph

Set 2

JointBayes Verification Graph






JointBayes Verification Graph

- -   uses large training set

Set 3

JointBayes Verification Graph

Set 3

JointBayes Verification Graph

Analysis of Rank-1 Identification for Varying Ages

The colors represent identification accuracy going from 0(=blue)–none of the true pairs were matched to 1(=red)–all possible combinations of probe and gallery were matched per probe and gallery ages