Identification Rate vs. Distractors Size


Algorithm Date Submitted Set 1 Set 2 Set 3 Data Set Size
Google - FaceNet v8 10/23/2015 74.594% 74.585% 74.558% Large
SIATMMLAB TencentVision 12/1/2016 71.247% 71.283% 71.256% Large
iBug (Reported by Author) 04/28/2017 66.94% Small
Yang Sun (Reported by Author) 06/05/2017 65.491% 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
SIAT_MMLAB 3/29/2016 55.304% 53.41% 53.049% Small
NTechLAB - facenx_large 10/20/2015 52.716% 52.733% 52.743% 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
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 04/7/2017 39.859% 39.949% 39.895% Large
Vocord - DeepVo1 08/3/2016 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
SIATMMALB_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.
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.
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.
DeepSense V2 Adopt a network of very deep ResNet CNNs, learned by combined supervision(identification loss(softmax loss), verification loss, triplet loss).

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
Google - FaceNet v8 10/23/2015 74.594% 74.585% 74.558% Large
SIATMMLAB TencentVision 12/1/2016 71.247% 71.283% 71.256% Large
iBug (Reported by Author) 04/28/2017 66.94% Small
Yang Sun (Reported by Author) 06/05/2017 65.491% 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
SIAT_MMLAB 3/29/2016 55.304% 53.41% 53.049% Small
NTechLAB - facenx_large 10/20/2015 52.716% 52.733% 52.743% 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
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 04/7/2017 39.859% 39.949% 39.895% Large
Vocord - DeepVo1 08/3/2016 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
SIATMMALB_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.
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.
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.
DeepSense V2 Adopt a network of very deep ResNet CNNs, learned by combined supervision(identification loss(softmax loss), verification loss, triplet loss).

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
SIATMMLAB TencentVision 12/1/2016 67.954% 67.954% 67.954% Large
Yang Sun (Reported by Author) 06/05/2017 63.623% Large
DeepSense V2 1/22/2017 56.767% 56.767% 56.767% 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
SIAT_MMLAB 3/29/2016 50.144% 51.155% 51.155% Small
NTechLAB - facenx_large 10/20/2015 45.381% 45.507% 44.37% Large
iBug (Reported by Author) 04/28/2017 44.947% Small
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
Vocord - DeepVo1 08/3/2016 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
DeepSense - Small 07/31/2016 29.61% 29.61% 29.61% Small
Progressor 09/13/2017 29.61% 29.61% 29.61% Large
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 04/7/2017 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
SIATMMALB_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.
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.
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.
DeepSense V2 Adopt a network of very deep ResNet CNNs, learned by combined supervision(identification loss(softmax loss), verification loss, triplet loss).

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