MultiFace: A Python repository from yuanwei0908

MultiFace_Pytorch

1. Intro

This repo is a implementation of MultiFace
Pretrained models are posted, include the MobileFacenet and IR-SE100 in the original paper

2. Pretrained Models & Performance

Loss	AgeDB-30(%)	CFP-FP(%)	calfw(%)	cplfw(%)
Multi-Arcface(N=2)	98.20	98.30	96.02	93.17
Multi-Cosface(N=2)	98.20	98.40	96.07	93.06

Mobilefacenet @ GoogleDrive

Loss	LFW(%)	CFP-FP(%)	AgeDB-30(%)	calfw(%)	cplfw(%)
Softmax	99.22	92.84	94.00	93.80	88.30
Multi-Softmax(N=4)	99.40	95.46	95.25	95.15	90.22

Loss	LFW(%)	CFP-FP(%)	AgeDB-30(%)	calfw(%)	cplfw(%)
Arcface	99.45	92.27	96.03	95.12	87.75
Multi-Arcface(N=2)	99.52	93.41	96.35	95.28	88.65
Cosface	99.43	92.83	95.77	94.97	88.88
Multi-Cosface(N=2)	99.50	93.58	96.17	95.20	89.03
Multi-Cosface(N=4)	99.60	94.11	96.13	95.18	89.47

3. Training and Testing

3.1 Prepare training dataset

download the refined dataset: (emore recommended, our method is suitable for larger dataset )

emore dataset @ BaiduDrive, emore dataset @ Dropbox

Note: If you use the refined MS1M dataset, please cite the original papers.
after unzip the files to 'data' path, run :
```
python prepare_data.py
```
after the execution, you should find following structure:

faces_emore/
            ---> agedb_30
            ---> calfw
            ---> cfp_ff
            --->  cfp_fp
            ---> cfp_fp
            ---> cplfw
            --->imgs
            ---> lfw
            ---> vgg2_fp

3.2 Training:

'''
# mobilefacenet loss:softmax  num_sphere:2 
python train.py --num_sphere 2 --work_path [save log and model information]

# mobilefacenet loss:arcface  num_sphere:2 
python train.py --num_sphere 2 --arcface_loss --work_path [save log and model information]

# mobilefacenet loss:cosface  num_sphere:2 
python train.py --num_sphere 2 --am_softmax_loss --work_path [save log and model information]

# ir-se100 loss:arcface  num_sphere:2 
python train.py --num_sphere 2 --arcface_loss --net ir_se -depth 100 --work_path [save log and model information]

# ir-se100 loss:cosface  num_sphere:2 
python train.py --num_sphere 2 --am_softmax_loss --net ir_se -depth 100 --work_path [save log and model information]

3.3 Testing

Evaluating the model on LFW, Age-DB, CFP-FP, CALFW, CPLFW

#pretrained mobilefacenet
python train.py --pretrain --pretrained_model_path [mobilefacenet_pretrained_model_path]

#pretrained ir-se 100
python train.py --pretrain -net ir_se -depth 100 --work_path [resnet_pretrained_model_path]

To evaluate on Megaface, please refer to megaface-evaluation.

4. References

This code is based on the implementations of TreB1eN/InsightFace_Pytorch and deepinsight/insightface

Contact

Email :xujing.may@gmail.com

yuanwei0908/MultiFace