FaceScrub Results
Identification Rate vs. Distractors Size
| 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 | 83.29% | 83.267% | 83.295% | Large |
| Yang Sun (Reported by Author) | 06/05/2017 | 81.326% | Large | ||
| DeepSense V2 | 1/22/2017 | 81.298% | 81.298% | 81.298% | Large |
| iBug (Reported by Author) | 04/28/2017 | 80.277% | Small | ||
| 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% | 77.99% | Small |
| GRCCV | 12/1/2016 | 77.677% | 77.021% | 77.147% | Small |
| XT-tech V2 | 08/30/2017 | 77.239% | 77.239% | 77.239% | Large |
| Beijing Faceall Co. - FaceAll V2 | 04/28/2017 | 76.661% | 76.643% | 76.607% | Small |
| SphereFace - Small | 12/1/2016 | 75.766% | 75.765% | 75.77% | Small |
| Vocord - DeepVo1 | 08/3/2016 | 75.127% | 75.093% | 75.125% | Large |
| DeepSense - Large | 07/31/2016 | 74.799% | 74.78% | 74.813% | Large |
| SIATMMLAB TencentVision | 12/1/2016 | 74.207% | 74.213% | 74.195% | Large |
| Shanghai Tech | 08/13/2016 | 74.049% | 74.032% | 74.02% | Large |
| CVTE | 08/30/2017 | 73.521% | 73.501% | 73.504% | Large |
| NTechLAB - facenx_large | 10/20/2015 | 73.3% | 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% | 58.21% | Small |
| 3DiVi Company - tdvm6 | 10/27/2015 | 33.705% | 33.69% | 33.667% | Small |
| Joint Bayes | 10/20/2015 | 3.021% | 3.223% | 3.245% | Small |
| LBP | 10/20/2015 | 2.326% | 2.32% | 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). |
- - uses large training set
Set 1
Set 1
- - uses large training set
Set 2
Set 2
- - uses large training set
Set 3
Set 3
Identification Rank vs. Rank
| 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 | 83.29% | 83.267% | 83.295% | Large |
| Yang Sun (Reported by Author) | 06/05/2017 | 81.326% | Large | ||
| DeepSense V2 | 1/22/2017 | 81.298% | 81.298% | 81.298% | Large |
| iBug (Reported by Author) | 04/28/2017 | 80.277% | Small | ||
| 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% | 77.99% | Small |
| GRCCV | 12/1/2016 | 77.677% | 77.021% | 77.147% | Small |
| XT-tech V2 | 08/30/2017 | 77.239% | 77.239% | 77.239% | Large |
| Beijing Faceall Co. - FaceAll V2 | 04/28/2017 | 76.661% | 76.643% | 76.607% | Small |
| SphereFace - Small | 12/1/2016 | 75.766% | 75.765% | 75.77% | Small |
| Vocord - DeepVo1 | 08/3/2016 | 75.127% | 75.093% | 75.125% | Large |
| DeepSense - Large | 07/31/2016 | 74.799% | 74.78% | 74.813% | Large |
| SIATMMLAB TencentVision | 12/1/2016 | 74.207% | 74.213% | 74.195% | Large |
| Shanghai Tech | 08/13/2016 | 74.049% | 74.032% | 74.02% | Large |
| CVTE | 08/30/2017 | 73.521% | 73.501% | 73.504% | Large |
| NTechLAB - facenx_large | 10/20/2015 | 73.3% | 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% | 58.21% | Small |
| 3DiVi Company - tdvm6 | 10/27/2015 | 33.705% | 33.69% | 33.667% | Small |
| Joint Bayes | 10/20/2015 | 3.021% | 3.223% | 3.245% | Small |
| LBP | 10/20/2015 | 2.326% | 2.32% | 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
- - 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
| Algorithm | Date Submitted | Set 1 | Set 2 | Set 3 | Data Set Size |
|---|---|---|---|---|---|
| DeepSense V2 | 1/22/2017 | 95.993% | 95.993% | 94.984% | Large |
| Yang Sun (Reported by Author) | 06/05/2017 | 95.178% | Large | ||
| Vocord - deepVo V3 | 04/27/2017 | 94.963% | 94.963% | 94.963% | Large |
| XT-tech V2 | 08/30/2017 | 93.695% | 93.695% | 93.695% | Large |
| iBug (Reported by Author) | 04/28/2017 | 92.639% | Small | ||
| YouTu Lab (Tencent Best-Image) | 04/08/2017 | 91.34% | 91.34% | 91.34% | 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 |
| CVTE | 08/30/2017 | 85.563% | 85.563% | 85.563% | 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 |
| Beijing Faceall Co. - FaceAll V2 | 04/28/2017 | 77.607% | 77.607% | 77.607% | Small |
| Vocord-deepVo1.2 | 12/1/2016 | 77.143% | 77.143% | 77.143% | Large |
| SIAT_MMLAB | 3/29/2016 | 76.72% | 76.72% | 76.72% | 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 |
| NTechLAB - facenx_small | 10/20/2015 | 66.366% | 66.427% | 66.427% | Small |
| Beijing Faceall Co. - FaceAll_1600 | 10/19/2015 | 63.96% | 64.983% | 63.96% | Large |
| Barebones_FR - cnn | 10/21/2015 | 59.036% | 59.036% | 59.036% | 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
- - 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 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.
