FGNet Results
Rank-1 Identification Accuracy with 1 Million Distractors (testing age-invariant recognition at scale)
| 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 | 66.940% | Small | ||
| Yang Sun (Reported by Author) | 06/05/2017 | 65.491% | 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 |
| DeepSense V2 | 1/22/2017 | 63.632% | 63.632% | 63.632% | Large |
| SIAT_MMLAB | 3/29/2016 | 55.304% | 53.410% | 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 |
| 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 | 43.540% | 43.576% | 43.522% | Small |
| 3DiVi Company - tdvm V2 | 04/15/2017 | 40.256% | 40.166% | 40.220% | 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% | 39.020% | Small |
| NTechLAB - facenx_small | 10/20/2015 | 29.168% | 29.150% | 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 | 25.020% | 25.045% | 24.973% | Large |
| GRCCV | 12/1/2016 | 24.783% | 24.802% | 24.783% | Small |
| 3DiVi Company - tdvm6 | 10/27/2015 | 15.780% | 15.825% | 15.770% | 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
- - uses large training set
Set 1
Set 1
- - uses large training set
Set 2
Set 2
- - uses large training set
Set 3
Set 3
Rank-1 Identification Accuracy with 1 Million Distractors (testing age-invariant recognition at scale)
| 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 | 66.940% | 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% | 53.410% | 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 |
| 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 | 43.540% | 43.576% | 43.522% | Small |
| 3DiVi Company - tdvm V2 | 04/15/2017 | 40.256% | 40.166% | 40.220% | 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% | 39.020% | Small |
| NTechLAB - facenx_small | 10/20/2015 | 29.168% | 29.150% | 29.168% | Small |
| Beijing Faceall Co. - FaceAll_1600 | 10/19/2015 | 26.155% | 26.155% | 26.137% | Large |
| Beijing Faceall Co. - FaceAll_Norm_1600 | 10/19/2015 | 25.020% | 25.045% | 24.973% | Large |
| Fudan University - FUDAN-CS_SDS | 1/29/2017 | 25.568% | 25.559% | 25.623% | Small |
| GRCCV | 12/1/2016 | 24.783% | 24.802% | 24.783% | Small |
| 3DiVi Company - tdvm6 | 10/27/2015 | 15.780% | 15.825% | 15.770% | Small |
| Barebones_FR - cnn | 10/21/2015 | 7.136% | 10.709% | 10.736% | Small |
Method Details
| Algorithm | Details (Provided by Team) |
|---|---|
| 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). |
- - uses large training set
Set 1
Set 1
- - uses large training set
Set 2
Set 2
- - uses large training set
Set 3
Set 3
Verification TAR for 10-6 FAR (testing age-invariant recognition at scale)
| Algorithm | Date Submitted | Set 1 | Set 2 | Set 3 | Data Set Size |
|---|---|---|---|---|---|
| Google - FaceNet v8 | 10/23/2015 | 75.550% | 75.550% | 75.550% | Large |
| SIATMMLAB TencentVision | 12/1/2016 | 67.954% | 67.954% | 67.954% | Small |
| Yang Sun (Reported by Author) | 06/05/2017 | 63.623% | Large | ||
| DeepSense V2 | 1/22/2017 | 56.767% | 56.767% | 56.767% | Large |
| Vocord - deepVo V3 | 04/27/2017 | 53.573% | 53.573% | 54.637% | Large |
| YouTu Lab (Tencent Best-Image) | 04/08/2017 | 53.681% | 53.681% | 53.681% | Large |
| SIAT_MMLAB | 3/29/2016 | 50.144% | 51.155% | 51.155% | Small |
| NTechLAB - facenx_large | 10/20/2015 | 45.381% | 45.507% | 44.370% | 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 |
| DeepSense - Small | 07/31/2016 | 29.610% | 29.610% | 29.610% | Small |
| DeepSense - Large | 07/31/2016 | 29.177% | 29.177% | 29.177% | Large |
| Beijing Faceall Co. - FaceAll V2 | 04/28/2017 | 26.540% | 26.540% | 26.540% | Small |
| Shanghai Tech | 08/13/2016 | 27.463% | 26.416% | 26.416% | Large |
| 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 |
| 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
- - uses large training set
Set 1
Set 1
- - uses large training set
Set 2
Set 2
- - uses large training set
Set 3
Set 3
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
