This is the official repository of the paper:

ContraFace: Sample-Guided Contrastive Learning for Deep Face Recognition

Evaluation Results

Methods(%)	Venue	LFW	AgeDB	CFP-FP	CALFW	CPLFW	Average
CosFace	CVPR2018	99.81	98.11	98.12	95.76	92.28	96.816
ArcFace	CVPR2019	99.83	98.28	98.27	95.45	92.08	96.782
AFRN	ICCV2019	99.85	95.35	95.56	96.30	93.48	96.108
MV-Softmax	AAAI2020	99.80	97.95	98.28	96.10	92.83	96.992
CurricularFace	CVPR2020	99.80	98.32	98.37	96.20	93.13	97.164
BroadFace	ECCV2020	99.85	98.38	98.63	96.20	93.17	97.246
SCF-ArcFace	CVPR2021	99.82	98.30	98.40	96.12	93.16	97.160
MagFace	CVPR2021	99.83	98.17	98.46	96.15	92.87	97.096
AdaFace	CVPR2022	99.82	98.05	98.49	96.08	93.53	97.194
ContraFace	AAAI2023	99.83	98.37	98.60	96.20	93.27	97.254

Model Training

We take MS1MV2 as the training dataset which can be downloaded from insightface (MS1M-ArcFace in DataZoo). Set the config.output and config.rec in the config/config.py

Our model is trained on 4 NVIDIA A100 GPUs for about 39 hours.

An economic alternative is trained on CASIA-WebFace with ResNet50 as the backbone.

The config.eval_step is set to the number of steps in an epoch so the model would be evaluated at the end of every epoch.

The log files could be find in the project and the model file can be downloaded here (with extraction code g5by).

You may train a new model with the command below.

export OMP_NUM_THREADS=4
CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 \
--node_rank=0 --master_addr="127.0.0.1" --master_port=1235 train.py

Face recognition evaluation

Evaluation on LFW, AgeDb-30, CPLFW, CALFW and CFP-FP:

1. download the data from their offical webpages or the bin files from insightface datasets
2. set the config.rec to dataset folder e.g. data/faces_emore
3. set the config.val_targets for list of the evaluation dataset
4. just use the CallBackVerification in utils/utils_callbacks.py to evaluate a model
5. the evaluation of all of the five models is less than 100 seconds as we make it parrallel

Evaluation on IJB and megaface

You may find the evaluation scripts at insightface/recognition/_evaluation_.

As the original IJB scripts don't support parameter-only saved pytorch models, we change the code and put them with megaface scripts in eval folder. Examples are located in two sub-folders separately.

The evaluation of IJB-B and IJB-C cost about 15 mins and 33 mins on one 2080-Ti GPU.

The evaluation of megaface needs their linux-only develop suit. It's really time-consuming (2+3 hours on one 2080-Ti) sincethe operation on 1M images.

Acknowledgement

GitHub - deepinsight/insightface: State-of-the-art 2D and 3D Face Analysis Project You may also find more useful information here, especially in their issue.

GitHub - fdbtrs/ElasticFace: Official repository for ElasticFace: Elastic Margin Loss for Deep Face Recognition

GitHub - MLNLP-World/SimBiber: A tool for simplifying bibtex with official info

GitHub - yuchenlin/rebiber: A simple tool to update bib entries with their official information (e.g., DBLP or the ACL anthology).

GitHub - MLNLP-World/Paper-Writing-Tips: Paper Writing Tips

License

This project is licensed under the terms of the Attribution-NonCommercial-ShareAlike 4.0 
International (CC BY-NC-SA 4.0) license.