Our pre-trained CFSMs could be a plug-in to any SoTA Face recognition model. Here, we provide a training example based on the ArcFace Pytorch repo.
- Prepare the training data and add the data path to the configuration file.
- Train the model:
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch \
--nproc_per_node=8 \
--nnodes=1 \
--node_rank=0 \
--master_addr="127.0.0.1" \
--master_port=12348 train.py configs/arcface_ms1mv2_r50_ours
-
Pre-trained face synthesis model
In order to train the synthesis model for face recognition, 70K images of faces from WiderFace are used as target data. The pre-trained model could be downloaded here.
-
Trained face recognition models
Method | Arch | Train Dataset | Link |
---|---|---|---|
ArcFace+CFSM | R50 | MS1MV2 | gdrive, baidudrive (wyiv) |
AdaFace+CFSM | R100 | WebFace12M | gdrive, baidudrive (hatp) |
-
IJB-B dataset
Method Arch Train Dataset TAR@FAR=0.001% TAR@FAR=0.01% TAR@FAR=0.1% Rank1 Rank5 ArcFace R50 MS1MV2 87.26 94.01 95.95 94.61 96.52 ArcFace+CFSM R50 MS1MV2 90.95 94.61 96.21 94.96 96.84 -
IJB-C dataset
Method Arch Train Dataset TAR@FAR=0.0001% TAR@FAR=0.001% TAR@FAR=0.01% Rank1 Rank5 ArcFace R50 MS1MV2 87.24 93.32 95.61 95.89 97.08 ArcFace+CFSM R50 MS1MV2 89.34 94.06 95.90 96.31 97.48 -
TinyFace dataset
Method Arch Train Dataset Rank1 Rank5 ArcFace R50 MS1MV2 62.21 66.85 ArcFace+CFSM R50 MS1MV2 63.01 68.21 AdaFace R100 WebFace12M 72.29 74.97 AdaFace+CFSM R100 WebFace12M 73.87 76.77 -
IJB-S dataset
Method | Arch | Train Dataset | V2S Rank1 | V2S Rank5 | V2S 1% | V2S 10% | V2B Rank1 | V2B Rank5 | V2B 1% | V2B 10% | V2V Rank1 | V2V Rank5 | V2V 1% | V2V 10% |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
ArcFace | R50 | MS1MV2 | 58.78 | 66.40 | 40.99 | 50.45 | 60.66 | 67.43 | 43.12 | 51.38 | 14.81 | 26.72 | 2.51 | 5.72 |
ArcFace+CFSM | R50 | MS1MV2 | 63.86 | 69.95 | 47.86 | 56.44 | 65.95 | 71.16 | 47.28 | 57.24 | 21.38 | 35.11 | 2.96 | 7.41 |
AdaFace | R100 | WebFace12M | 71.35 | 76.24 | 59.40 | 66.34 | 71.93 | 76.56 | 59.37 | 66.68 | 36.71 | 50.03 | 4.62 | 11.84 |
AdaFace+CFSM | R100 | WebFace12M | 72.54 | 77.59 | 60.94 | 66.02 | 72.65 | 78.18 | 60.26 | 65.88 | 39.14 | 50.91 | 5.05 | 13.17 |