junuke/UNPG

UNPG

Unified Negative Pair Generation toward Well-discriminative Feature Space for Face Recognition

Data prepration

MS1MV2

MS1M-ArcFace (85K ids/5.8M images) download link

#Preprocess 'train.rec' and 'train.idx' to 'jpg'

# example
cd detection

python rec2image.py --include '{path}/face_emore' --output '{path}/MS1MV2'

K-FACE

K-FACE download link

"""
    ###################################################################

    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
    
    ###################################################################
"""

# example
cd detection

python align_kfaces.py --ori_data_path '/data/FACE/KFACE/High' --detected_data_path 'kface_retina_align_112x112'

IJBB & IJBC

download link

Please apply for permissions from NIST before your usage.

Evaluation

Performance on public benchmark datasets with ResNet-100

Method	IJB-B (1e-5)	IJB-B (1e-4)	IJB-C (1e-5)	IJB-C (1e-4)	MegaFace (Rank-1 acc)	LFW	Path
Circle-loss*	-	-	-	93.95	98.50	99.73	-
ArcFace*	-	94.20	-	95.60	98.35	99.82	-
MagFace*	90.36	94.51	94.08	95.97	-	99.83	-
CosFace	89.38	94.39	94.42	96.35	99.08	99.83	-
CosFace+UNPG	90.61	94.99	94.48	96.39	99.27	99.81	link
ArcFace	89.99	94.89	93.93	96.25	98.56	99.83	-
ArcFace+UNPG	90.57	95.04	94.47	96.33	98.82	99.83	link
MagFace	89.03	93.99	93.30	95.54	98.51	99.81	-
MagFace+UNPG	90.93	95.21	94.70	96.38	98.03	99.81	link

“*” indicates results from the original paper.

Example script

cd recognition

# example
python evaluation.py --weights 'face.r100.cos.unpg.wisk1.5.pt' --data 'ijbc.yaml' 
# --data (e.g., 'ijbb.yaml', 'bins.yaml')

Performance on K-FACE test datasets (Q1-Q4) with ResNet-34

Method	Q4 (1e-5)	Q4 (1e-4)	Q3 (1e-5)	Q3 (1e-4)	Q2 (1e-5)	Q2 (1e-4)	Q1 (1e-3)	Q1 (1e-2)	Path
ArcFace	0.05	0.29	2.06	4.40	26.56	41.29	94.00	100	-
SN-pair	3.50	7.21	17.67	21.16	21.93	33.26	91.80	97.60	-
MS-loss	5.68	8.70	15.15	18.74	38.33	46.64	94.60	99.20	-
MixFace	7.11	10.92	9.19	22.55	39.09	44.48	97.00	100	-
Circle-loss	18.08	25.05	33.56	41.54	71.38	77.93	100	100	-
Arc+UNPG	29.89	50.43	51.59	60.88	91.28	93.26	100	100	link

cd recognition

# example
python evaluation.py --weights 'kface.r34.arc.unpg.wisk1.0.pt' --data 'kface.yaml' 

Training

Example script (FACE)

cd recognition

# example 
python train.py --model 'iresnet-100' --head 'arcface' --aux 'unpg' --data 'data/face.yaml' --hyp 'data/hyp.yaml' --name 'example' --device 0,1

Example script (KFACE)

cd recognition

# example 
python train.py --model 'iresnet-34' --head 'arcface' --aux 'unpg' --data 'data/kface.yaml' --hyp 'data/hyp.yaml' --name 'example' --device 0,1

License

The source code in this repository is released under the Apache License 2.0.

The Apache License 2.0 does not apply to any model weights, including:

• Pretrained UNPG models (e.g., face.r100.cos.unpg.wisk1.5.pt, kface.r34.arc.unpg.wisk1.0.pt). These were trained on MS1MV2 (distributed by InsightFace) and K-FACE, whose terms permit non-commercial research use only. Accordingly, these weights are provided for non-commercial research purposes only.
• detection/retinaface/weights/mobilenet0.25_Final.pth, a third-party model included for convenience. It is subject to its original license and training-data terms, not this repository's license.

If you wish to use the models commercially, you must obtain the appropriate permissions from the respective dataset and model rights holders.