Rank-1 Identification Accuracy with 1 Million Distractors

Large is >500K photos trained.
Algorithm Date Submitted Set 1 Set 2 Set 3 Data Set Size
Vocord - deepVo V3 04/27/2017 91.763% 91.711% 91.704% Large
YouTu Lab (Tencent Best-Image) 04/08/2017 83.290% 83.267% 83.295% Large
DeepSense V2 1/22/2017 81.298% 81.298% 81.298% Large
Vocord-deepVo1.2 12/1/2016 80.258% 80.195% 80.241% Large
Fudan University - FUDAN-CS_SDS 1/29/2017 77.982% 78.006% 77.990% Small
GRCCV 12/1/2016 77.677% 77.021% 77.147% Small
Beijing Faceall Co. - FaceAll V2 04/28/2017 76.661% 76.643% 76.607% Small
SphereFace - Small 12/1/2016 75.766% 75.765% 75.770% Small
Vocord - DeepVo1 08/3/2016 75.127% 75.093% 75.125% Large
DeepSense - Large 07/31/2016 74.799% 74.780% 74.813% Large
SIATMMLAB TencentVision 12/1/2016 74.207% 74.213% 74.195% Large
Shanghai Tech 08/13/2016 74.049% 74.032% 74.020% Large
NTechLAB - facenx_large 10/20/2015 73.300% 73.309% 73.287% Large
ForceInfo 04/7/2017 72.11% 72.084% 72.121% Large
3DiVi Company - tdvm V2 04/15/2017 71.742% 71.727% 71.703% Large
DeepSense - Small 07/31/2016 70.983% 70.948% 70.962% Small
Google - FaceNet v8 10/23/2015 70.496% 70.492% 70.551% Large
SIAT_MMLAB 3/29/2016 65.233% 65.223% 65.229% Small
Beijing Faceall Co. - FaceAll_Norm_1600 10/19/2015 64.803% 64.798% 64.826% Large
Beijing Faceall Co. - FaceAll_1600 10/19/2015 63.977% 63.962% 63.993% Large
Barebones_FR - cnn 10/21/2015 59.363% 59.379% 59.389% Small
NTechLAB - facenx_small 10/20/2015 58.218% 58.208% 58.210% Small
3DiVi Company - tdvm6 10/27/2015 33.705% 33.690% 33.667% Small
Joint Bayes 10/20/2015 3.021% 3.223% 3.245% Small
LBP 10/20/2015 2.326% 2.320% 2.318% 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

Rank-1 Identification Accuracy with 1 Million Distractors

Large is >500K photos trained.
Algorithm Date Submitted Set 1 Set 2 Set 3 Data Set Size
Vocord - deepVo V3 04/27/2017 91.763% 91.711% 91.704% Large
YouTu Lab (Tencent Best-Image) 04/08/2017 83.290% 83.267% 83.295% Large
DeepSense V2 1/22/2017 81.298% 81.298% 81.298% Large
Vocord-deepVo1.2 12/1/2016 80.258% 80.195% 80.241% Large
Fudan University - FUDAN-CS_SDS 1/29/2017 77.982% 78.006% 77.990% Small
GRCCV 12/1/2016 77.677% 77.021% 77.147% Small
Beijing Faceall Co. - FaceAll V2 04/28/2017 76.661% 76.643% 76.607% Small
SphereFace - Small 12/1/2016 75.766% 75.765% 75.770% Small
Vocord - DeepVo1 08/3/2016 75.127% 75.093% 75.125% Large
DeepSense - Large 07/31/2016 74.799% 74.780% 74.813% Large
SIATMMLAB TencentVision 12/1/2016 74.207% 74.213% 74.195% Large
Shanghai Tech 08/13/2016 74.049% 74.032% 74.020% Large
NTechLAB - facenx_large 10/20/2015 73.300% 73.309% 73.287% Large
ForceInfo 04/7/2017 72.11% 72.084% 72.121% Large
3DiVi Company - tdvm V2 04/15/2017 71.742% 71.727% 71.703% Large
DeepSense - Small 07/31/2016 70.983% 70.948% 70.962% Small
Google - FaceNet v8 10/23/2015 70.496% 70.492% 70.551% Large
SIAT_MMLAB 3/29/2016 65.233% 65.223% 65.229% Small
Beijing Faceall Co. - FaceAll_Norm_1600 10/19/2015 64.803% 64.798% 64.826% Large
Beijing Faceall Co. - FaceAll_1600 10/19/2015 63.977% 63.962% 63.993% Large
Barebones_FR - cnn 10/21/2015 59.363% 59.379% 59.389% Small
NTechLAB - facenx_small 10/20/2015 58.218% 58.208% 58.210% Small
3DiVi Company - tdvm6 10/27/2015 33.705% 33.690% 33.667% Small
Joint Bayes 10/20/2015 3.021% 3.223% 3.245% Small
LBP 10/20/2015 2.326% 2.320% 2.318% 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

Verification TAR for 10-6 FAR

Large is >500K photos trained.
Algorithm Date Submitted Set 1 Set 2 Set 3 Data Set Size
DeepSense V2 1/22/2017 95.993% 95.993% 94.984% Large
Vocord - deepVo V3 04/27/2017 94.963% 94.963% 94.963% Large
YouTu Lab (Tencent Best-Image) 04/08/2017 91.340% 91.340% 91.340% Large
SphereFace - Small 12/1/2016 90.045% 89.355% 90.045% Small
DeepSense - Large 07/31/2016 87.764% 87.764% 87.764% Large
SIATMMLAB TencentVision 12/1/2016 87.272% 87.021% 87.021% Large
Google - FaceNet v8 10/23/2015 86.473% 86.386% 86.473% Large
Shanghai Tech 08/13/2016 86.369% 86.369% 86.369% Large
ForceInfo 04/7/2017 85.918% 85.619% 85.619% Large
3DiVi Company - tdvm V2 04/15/2017 85.411% 85.411% 85.411% Large
NTechLAB - facenx_large 10/20/2015 85.081% 85.081% 85.081% Large
DeepSense - Small 07/31/2016 82.851% 82.851% 82.851% Small
Fudan University - FUDAN-CS_SDS 1/29/2017 79.199% 79.199% 79.199% Small
Vocord-deepVo1.2 12/1/2016 77.143% 77.143% 77.143% Large
SIAT_MMLAB 3/29/2016 76.720% 76.720% 76.720% Small
Beijing Faceall Co. - FaceAll V2 04/28/2017 77.607% 77.607% 77.607% Small
GRCCV 12/1/2016 74.887% 74.887% 75.918% Small
Vocord - DeepVo1 08/3/2016 67.318% 67.318% 67.318% Large
Beijing Faceall Co. - FaceAll_Norm_1600 10/19/2015 67.118% 67.118% 67.118% Large
Beijing Faceall Co. - FaceAll_1600 10/19/2015 63.960% 64.983% 63.960% Large
Barebones_FR - cnn 10/21/2015 59.036% 59.036% 59.036% Small
NTechLAB - facenx_small 10/20/2015 66.366% 66.427% 66.427% Small
3DiVi Company - tdvm6 10/27/2015 36.927% 37.967% 36.927% Small
Joint Bayes 10/20/2015 2.173% 2.204% 2.204% Small
LBP 10/20/2015 1.465% 1.465% 1.465% 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 Poses

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. White color indicates combinations of poses that did not exist in our test set.