Installation

Setup with Conda

# create a new environment
conda create --name insightKface python=3.7 # or over
conda activate insightKface
#install the appropriate cuda version of pytorch(https://pytorch.org/)#example:
conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c conda-forge
# install requirements
pip install -r requirements.txt

Data prepration

K-FACE Database

K-FACE AI-hub.

Detail configuration about K-FACE is provided in the paper below.

K-FACE: A Large-Scale KIST Face Database in Consideration with Unconstrained Environments

K-FACE sample images

Structure of the K-FACE database

Configuration of K-FACE

Detection & Alignment on K-FACE

""" ################################################################### K-Face : Korean Facial Image AI Dataset url : http://www.aihub.or.kr/aidata/73 Directory structure : High-ID-Accessories-Lux-Emotion ID example : '19062421' ... '19101513' len 400 Accessories example : 'S001', 'S002' .. 'S006' len 6 Lux example : 'L1', 'L2' .. 'L30' len 30 Emotion example : 'E01', 'E02', 'E03' len 3 ###################################################################"""# examplecd detection
python align_kfaces.py --ori_data_path '/data/FACE/KFACE/High' --detected_data_path 'kface_retina_align_112x112'

Training and test datasets on K-FACE

Train ID	Accessories	Lux	Expression	Pose	#Image	Variance
T1	A1	1000	E1	C4-10	2,590	Very Low
T2	A1-2	400-1000	E1	C4-10	46,620	Low
T3	A1-A4	200-1000	E1-2	C4-13	654,160	Middle
T4	A1-A6	40-1000	E1-3	C1-20	3,862,800	High
Test ID	Accessories	Lux	Expression	Pose	#Pairs	Variance
Q1	A1	1000	E1	C4-10	1,000	Very Low
Q2	A1-2	400-1000	E1	C4-10	100,000	Low
Q3	A1-4	200-1000	E1-2	C4-13	100,000	Middle
Q4	A1-6	40-1000	E1-3	C1-20	100,000	High

MS1M-RetinaFace (MS1M-R)

MS1M-RetinaFace download link:

1. The Lightweight Face Recognition Challenge & Workshop.

2. https://github.com/deepinsight/insightface/wiki/Dataset-Zoo

#Preprocess 'train.rec' and 'train.idx' to 'jpg'# examplecd detection
python rec2image.py --include '/data/FACE/ms1m-retinaface-t1/' --output 'MS1M-RetinaFace'

Inference

After downloading the pretrained model, run test.py.

Pretrained Model

For all experiments, ResNet-34 was chosen as the baseline backbone.

The model was trained on KFACE

Head&Loss	Q1	Q2	Q3	Q4
ArcFace (s=16, m=0.25)	98.30	94.77	87.87	85.41
SN-pair (s=64)	99.20	95.01	91.84	89.74
MixFace (e=1e-22, m=0.25)	100	96.37	92.36	89.80

Note:

• For ArcFace, We tested (s,m)={(16,0.5), (32,0.25), (64,0.25), (32,0.5), (64,0.5)}, but the model was not trained properly So, we apply (s,m)=(16,0.25).

cd recognition
# example
python test.py --weights 'kface.mixface.1e-22m0.25.best.pt' --dataset 'kface' --data_cfg 'data/KFACE/kface.T4.yaml'

The model was trained on MS1M-R

Head&Loss	Q2	Q3	Q4	LFW	CFP-FP	AgeDB-30
ArcFace (s=64, m=0.5)	98.71	86.60	82.03	99.80	98.41	98.80
SN-pair (s=64)	92.85	76.36	70.08	99.55	96.20	95.46
MixFace (e=1e-22, m=0.5)	97.36	82.89	76.95	99.68	97.74	97.25

cd recognition
# example
python test.py --weights 'face.mixface.1e-22m0.5.best.pt' --dataset 'face' --data_cfg 'data/face.all.yaml'

The model was trained on MS1M-R+T4

Head&Loss	Q2	Q3	Q4	LFW	CFP-FP	AgeDB-30
ArcFace (s=8, m=0.25)	76.58	73.13	71.38	99.46	96.75	93.83
SN-pair (s=64)	98.37	94.98	93.33	99.45	94.90	93.45
MixFace (e=1e-22, m=0.5)	99.27	96.85	94.79	99.53	96.32	95.56

Note:

• For ArcFace, we tested (s,m)={(8, 0.5), (16, 0.25), (16,0.5), (32,0.25), (64,0.25), (32,0.5), (64,0.5)}, but the model was not trained properly So, we apply (s,m)=(8,0.25).

cd recognition
# example
python test.py --weights 'merge.mixface.1e-22m0.5.best.pt' --dataset 'merge' --data_cfg 'data/merge.yaml'

Training

Multi-GPU DataParallel Mode

Example script for training on KFACE

cd recognition
# example 
python train.py --dataset 'kface' --head 'mixface' --data_cfg 'data/KFACE/kface.T4.yaml' --hyp 'data/face.hyp.yaml' --head_cfg 'models/head.kface.cfg.yaml' --name 'example' --device 0,1

Multi-GPU DistributedDataParallel Mode

Example script for training on KFACE

cd recognition
# example
python -m torch.distributed.launch --nproc_per_node 2 train.py --dataset 'kface' --head 'mixface' --data_cfg 'data/KFACE/kface.T4.yaml' --hyp 'data/face.hyp.yaml' --head_cfg 'models/head.kface.cfg.yaml' --name 'example' --device 0,1

Note:

• For MS1M-R, change args --dataset face, --data_cfg data/face.all.yaml, and --head_cfg model/head.face.cfg.yaml.
• For MS1M-R+T4, change args --dataset merge, --data_cfg data/merge.yaml, and --head_cfg model/head.merge.cfg.yaml.
• The args --nodrop should be used if you train with the metric loss(e.g., SN-pair, N-pair, etc.) on MS1M-R or MS1M-R+T4.
• The args --double should be used if you train with the metric loss(e.g., SN-pair, N-pair, etc.) or MixFace on MS1M-R or MS1M-R+T4.
• DistributedDataParallel is only available to classification loss(e.g., arcface, cosface, etc.)

Reference code

Thanks for these source codes porviding me with knowledges to complete this repository.

1. https://github.com/biubug6/Pytorch_Retinaface.
2. https://github.com/deepinsight/insightface.
3. https://github.com/ultralytics/yolov5