Juruobudong/AdaFace

AdaFace: Quality Adaptive Margin for Face Recognition

Official github repository for AdaFace: Quality Adaptive Margin for Face Recognition. The paper (https://arxiv.org/abs/2204.00964) is presented in CVPR 2022 (Oral).

Abstract: Recognition in low quality face datasets is challenging because facial attributes are obscured and degraded. Advances in margin-based loss functions have resulted in enhanced discriminability of faces in the embedding space. Further, previous studies have studied the effect of adaptive losses to assign more importance to misclassified (hard) examples. In this work, we introduce another aspect of adaptiveness in the loss function, namely the image quality. We argue that the strategy to emphasize misclassified samples should be adjusted according to their image quality. Specifically, the relative importance of easy and hard samples should be based on the sample's image quality. We propose a new loss function that emphasizes samples of different difficulty based on their image quality. Our method achieves this in the form of an adaptive margin function by approximating the image quality with feature norms. Extensive experiments show that our method, AdaFace, improves the face recognition performance over the state-of-the-art (SoTA) on four datasets (IJB-B, IJB-C, IJB-S and TinyFace).

@inproceedings{kim2022adaface,
  title={AdaFace: Quality Adaptive Margin for Face Recognition},
  author={Kim, Minchul and Jain, Anil K and Liu, Xiaoming},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2022}
}

Demo Comparison between AdaFace and ArcFace on Low Quality Images

The demo shows a comparison between AdaFace and ArcFace on a live video. To show how model performs with low quality images, we show original, blur+ and blur++ setting where blur++ means it is heavily blurred. The numbers with colorbox show the cosine similarity between the live image and the cloest matching gallery image. The statistics on the bottom show the cumulative count of true positive match for blur++ setting. AdaFace has high true positive rate. It also shows it is less prone to making false positive (red) mistakes as sometimes observed in ArcFace.

Usage

import torch
from head import AdaFace

# typical inputs with 512 dimension
B = 5
embbedings = torch.randn((B, 512)).float()  # latent code
norms = torch.norm(embbedings, 2, -1, keepdim=True)
normalized_embedding  = embbedings / norms
labels =  torch.randint(70722, (B,))

# instantiate AdaFace
adaface = AdaFace(embedding_size=512,
                  classnum=70722,
                  m=0.4,
                  h=0.333,
                  s=64.,
                  t_alpha=0.01,)

# calculate loss
cosine_with_margin = adaface(normalized_embedding, norms, labels)
loss = torch.nn.CrossEntropyLoss()(cosine_with_margin, labels)

Installation

conda create --name adaface pytorch==1.8.0 torchvision==0.9.0 cudatoolkit=10.2 -c pytorch
conda activate adaface
conda install scikit-image matplotlib pandas scikit-learn 
pip install -r requirements.txt

Dataset (MS1MV2)

1. Download MS1M-ArcFace (85K ids/5.8M images) from InsightFace and unzip at DATASET_ROOT
2. Unpack mxrecord files to imgs with the following code.

python convert.py --rec_path <DATASET_ROOT>/faces_emore

Train

# training small model (resnet18) on a subset of MS1MV2 dataset
python main.py \
    --data_root <DATASET_ROOT> \
    --train_data_path faces_emore/imgs \
    --val_data_path faces_emore \
    --train_data_subset \
    --prefix run_ir18_ms1mv2_subset \
    --gpus 2 \
    --use_16bit \
    --batch_size 512 \
    --num_workers 16 \
    --epochs 26 \
    --lr_milestones 12,20,24 \
    --head adaface \
    --m 0.4 \
    --h 0.333 \
    --low_res_augmentation_prob 0.2 \
    --crop_augmentation_prob 0.2 \
    --photometric_augmentation_prob 0.2

Pretrained Models

Arch	Dataset	Link
R18	CASIA-WebFace	gdrive
R18	VGGFace2	gdrive
R18	WebFace4M	gdrive
R50	CASIA-WebFace	gdrive
R50	WebFace4M	gdrive
R50	MS1MV2	gdrive
R100	MS1MV2	gdrive
R100	MS1MV3	gdrive
R100	WebFace4M	gdrive
R100	WebFace12M	gdrive

Inferece

Download the pretrained adaface model and place it in pretrained/

For inference, refer to

python inference.py

We provide example images for inference.

img1	img2	img3

The similarity score result should be

tensor([[ 1.0000,  0.7334, -0.0655],
        [ 0.7334,  1.0000, -0.0277],
        [-0.0655, -0.0277,  1.0000]], grad_fn=<MmBackward0>)

Validation

High Quality Image Validation Sets (LFW, CFPFP, CPLFW, CALFW, AGEDB)

For evaluation on 5 HQ image validation sets with pretrained models, refer to

bash validation_hq/eval_5valsets.sh

Arch	Dataset	Method	LFW	CFPFP	CPLFW	CALFW	AGEDB	AVG
R18	CASIA-WebFace	AdaFace	0.9913	0.9259	0.8700	0.9265	0.9272	0.9282
R18	VGGFace2	AdaFace	0.9947	0.9713	0.9172	0.9390	0.9407	0.9526
R18	WebFace4M	AdaFace	0.9953	0.9726	0.9228	0.9552	0.9647	0.9621
R50	CASIA-WebFace	AdaFace	0.9942	0.9641	0.8997	0.9323	0.9438	0.9468
R50	MS1MV2	AdaFace	0.9982	0.9786	0.9283	0.9607	0.9785	0.9688
R50	WebFace4M	AdaFace	0.9978	0.9897	0.9417	0.9598	0.9778	0.9734
R100	MS1MV2	AdaFace	0.9982	0.9849	0.9353	0.9608	0.9805	0.9719
R100	MS1MV3	AdaFace	0.9978	0.9891	0.9393	0.9602	0.9817	0.9736
R100	WebFace4M	AdaFace	0.9980	0.9917	0.9463	0.9605	0.9790	0.9751
R100	WebFace12M	AdaFace	0.9982	0.9926	0.9457	0.9612	0.9800	0.9755

Comparison with Other Methods

Arch	Dataset	Method	LFW	CFPFP	CPLFW	CALFW	AGEDB	AVG
R50	CASIA-WebFace	AdaFace	0.9942	0.9641	0.8997	0.9323	0.9438	0.9468
R50	CASIA-WebFace	(ArcFace)	0.9945	0.9521	NA	NA	0.9490	NA
R100	MS1MV2	AdaFace	0.9982	0.9849	0.9353	0.9608	0.9805	0.9719
R100	MS1MV2	(ArcFace)	0.9982	NA	0.9208	0.9545	NA	NA

Mixed Quality Scenario (IJBB, IJBC Dataset)

For IJBB, IJBC validation, refer to

cd validation_mixed
bash eval_ijb.sh

Arch	Dataset	Method	IJBB TAR@FAR=0.01%	IJBC TAR@FAR=0.01%
R18	VGG2	AdaFace	90.67	92.95
R18	WebFace4M	AdaFace	93.03	94.99
R50	WebFace4M	AdaFace	95.44	96.98
R50	MS1MV2	AdaFace	94.82	96.27
R100	MS1MV2	AdaFace	95.67	96.89
R100	MS1MV3	AdaFace	95.84	97.09
R100	WebFace4M	AdaFace	96.03	97.39
R100	WebFace12M	AdaFace	96.41	97.66

Comparison with Other Methods

• Numbers for other methods come from their respective papers.

Arch	Dataset	Method	Venue	IJBB TAR@FAR=0.01%	IJBC TAR@FAR=0.01%
R100	MS1MV2	AdaFace	CVPR22	95.67	96.89
R100	MS1MV2	(MagFace)	CVPR21	94.51	95.97
R100	MS1MV2	(SCF-ArcFace)	CVPR21	94.74	96.09
R100	MS1MV2	(BroadFace)	ECCV20	94.97	96.38
R100	MS1MV2	(CurricularFace)	CVPR20	94.80	96.10
R100	MS1MV2	(MV-Softmax)	AAAI20	93.60	95.20
R100	MS1MV2	(AFRN)	ICCV19	88.50	93.00
R100	MS1MV2	(ArcFace)	CVPR19	94.25	96.03
R100	MS1MV2	(CosFace)	CVPR18	94.80	96.37

Arch	Dataset	Method	IJBC TAR@FAR=0.01%
R100	WebFace4M	AdaFace	97.39
R100	WebFace4M	(CosFace)	96.86
R100	WebFace4M	(ArcFace)	96.77
R100	WebFace4M	(CurricularFace)	97.02

Arch	Dataset	Method	IJBC TAR@FAR=0.01%
R100	WebFace12M	AdaFace	97.66
R100	WebFace12M	(CosFace)	97.41
R100	WebFace12M	(ArcFace)	97.47
R100	WebFace12M	(CurricularFace)	97.51

Low Quality Scenario (IJBS)

For IJBB, IJBC validation, refer to

cd validation_lq
python validate_IJB_S.py

Comparison with Other Methods

			Sur-to-Single			Sur-to-Book			Sur-to-Sur			TinyFace
Arch	Method	Dataset	Rank1	Rank5	1%	Rank1	Rank5	1%	Rank1	Rank5	1%	rank1	rank5
R100	AdaFace	WebFace4M	70.42	75.29	58.27	70.93	76.11	58.02	35.05	48.22	4.96	72.02	74.52
R100	AdaFace	MS1MV2	65.26	70.53	51.66	66.27	71.61	50.87	23.74	37.47	2.50	68.21	71.54
R100	(CurricularFace)	MS1MV2	62.43	68.68	47.68	63.81	69.74	47.57	19.54	32.80	2.53	63.68	67.65
R100	(URL)	MS1MV2	58.94	65.48	37.57	61.98	67.12	42.73	NA	NA	NA	63.89	68.67
R100	(ArcFace)	MS1MV2	57.35	64.42	41.85	57.36	64.95	41.23	NA	NA	NA	NA	NA
R100	(PFE)	MS1MV2	50.16	58.33	31.88	53.60	61.75	35.99	9.20	20.82	0.84	NA	NA

• Sur-to-Single: Protocol comparing surveillance video (probe) to single enrollment image (gallery)
• Sur-to-Book: Protocol comparing surveillance video (probe) to all enrollment images (gallery)
• Sur-to-Sur: Protocol comparing surveillance video (probe) to surveillance video (gallery)