By Jia Guo and Jiankang Deng
The code of InsightFace is released under the MIT License. There is no limitation for both acadmic and commercial usage.
The training data containing the annotation (and the models trained with these data) are available for non-commercial research purposes only.
Please click the image to watch the Youtube video. For Bilibili users, click here.
2020.02.21: Instant discussion group created on QQ with group-id: 711302608. For English developers, see install tutorial here.
2020.02.16: RetinaFace now can detect faces with mask, for anti-CoVID19, see detail here
2019.08.10: We achieved 2nd place at WIDER Face Detection Challenge 2019.
2019.05.30: Presentation at cvmart
2019.04.30: Our Face detector (RetinaFace) obtains state-of-the-art results on the WiderFace dataset.
2019.04.14: We will launch a Light-weight Face Recognition challenge/workshop on ICCV 2019.
2019.04.04: Arcface achieved state-of-the-art performance (7/109) on the NIST Face Recognition Vendor Test (FRVT) (1:1 verification)
report (name: Imperial-000 and Imperial-001). Our solution is based on [MS1MV2+DeepGlintAsian, ResNet100, ArcFace loss].
2019.02.08: Please check https://github.com/deepinsight/insightface/tree/master/recognition for our parallel training code which can easily and efficiently support one million identities on a single machine (8* 1080ti).
2018.12.13: Inference acceleration TVM-Benchmark.
2018.10.28: Light-weight attribute model Gender-Age. About 1MB, 10ms on single CPU core. Gender accuracy 96% on validation set and 4.1 age MAE.
2018.10.16: We achieved state-of-the-art performance on Trillionpairs (name: nttstar) and IQIYI_VID (name: WitcheR).
- Introduction
- Training Data
- Train
- Pretrained Models
- Verification Results On Combined Margin
- Test on MegaFace
- 512-D Feature Embedding
- Third-party Re-implementation
In this repository, we provide training data, network settings and loss designs for deep face recognition. The training data includes the normalised MS1M, VGG2 and CASIA-Webface datasets, which were already packed in MXNet binary format. The network backbones include ResNet, MobilefaceNet, MobileNet, InceptionResNet_v2, DenseNet, DPN. The loss functions include Softmax, SphereFace, CosineFace, ArcFace and Triplet (Euclidean/Angular) Loss.
Our method, ArcFace, was initially described in an arXiv technical report. By using this repository, you can simply achieve LFW 99.80%+ and Megaface 98%+ by a single model. This repository can help researcher/engineer to develop deep face recognition algorithms quickly by only two steps: download the binary dataset and run the training script.
All face images are aligned by MTCNN and cropped to 112x112:
Please check Dataset-Zoo for detail information and dataset downloading.
- Please check src/data/face2rec2.py on how to build a binary face dataset. Any public available MTCNN can be used to align the faces, and the performance should not change. We will improve the face normalisation step by full pose alignment methods recently.
- Install
MXNetwith GPU support (Python 2.7).
pip install mxnet-cu90
- Clone the InsightFace repository. We call the directory insightface as
INSIGHTFACE_ROOT.
git clone --recursive https://github.com/deepinsight/insightface.git
- Download the training set (
MS1M-Arcface) and place it in$INSIGHTFACE_ROOT/datasets/. Each training dataset includes at least following 6 files:
faces_emore/
train.idx
train.rec
property
lfw.bin
cfp_fp.bin
agedb_30.binThe first three files are the training dataset while the last three files are verification sets.
- Train deep face recognition models.
In this part, we assume you are in the directory
$INSIGHTFACE_ROOT/recognition/.
export MXNET_CPU_WORKER_NTHREADS=24
export MXNET_ENGINE_TYPE=ThreadedEnginePerDevicePlace and edit config file:
cp sample_config.py config.py
vim config.py # edit dataset path etc..We give some examples below. Our experiments were conducted on the Tesla P40 GPU.
(1). Train ArcFace with LResNet100E-IR.
CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network r100 --loss arcface --dataset emoreIt will output verification results of LFW, CFP-FP and AgeDB-30 every 2000 batches. You can check all options in config.py. This model can achieve LFW 99.80+ and MegaFace 98.3%+.
(2). Train CosineFace with LResNet50E-IR.
CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network r50 --loss cosface --dataset emore(3). Train Softmax with LMobileNet-GAP.
CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network m1 --loss softmax --dataset emore(4). Fine-turn the above Softmax model with Triplet loss.
CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --network m1 --loss triplet --lr 0.005 --pretrained ./models/m1-softmax-emore,1- Verification results.
LResNet100E-IR network trained on MS1M-Arcface dataset with ArcFace loss:
| Method | LFW(%) | CFP-FP(%) | AgeDB-30(%) |
|---|---|---|---|
| Ours | 99.80+ | 98.0+ | 98.20+ |
You can use $INSIGHTFACE/src/eval/verification.py to test all the pre-trained models.
Please check Model-Zoo for more pretrained models.
A combined margin method was proposed as a function of target logits value and original θ:
COM(θ) = cos(m_1*θ+m_2) - m_3
For training with m1=1.0, m2=0.3, m3=0.2, run following command:
CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train_softmax.py --network r100 --loss combined --dataset emore
Results by using MS1M-IBUG(MS1M-V1)
| Method | m1 | m2 | m3 | LFW | CFP-FP | AgeDB-30 |
|---|---|---|---|---|---|---|
| W&F Norm Softmax | 1 | 0 | 0 | 99.28 | 88.50 | 95.13 |
| SphereFace | 1.5 | 0 | 0 | 99.76 | 94.17 | 97.30 |
| CosineFace | 1 | 0 | 0.35 | 99.80 | 94.4 | 97.91 |
| ArcFace | 1 | 0.5 | 0 | 99.83 | 94.04 | 98.08 |
| Combined Margin | 1.2 | 0.4 | 0 | 99.80 | 94.08 | 98.05 |
| Combined Margin | 1.1 | 0 | 0.35 | 99.81 | 94.50 | 98.08 |
| Combined Margin | 1 | 0.3 | 0.2 | 99.83 | 94.51 | 98.13 |
| Combined Margin | 0.9 | 0.4 | 0.15 | 99.83 | 94.20 | 98.16 |
Please check $INSIGHTFACE_ROOT/Evaluation/megaface/ to evaluate the model accuracy on Megaface. All aligned images were already provided.
In this part, we assume you are in the directory $INSIGHTFACE_ROOT/deploy/. The input face image should be generally centre cropped. We use RNet+ONet of MTCNN to further align the image before sending it to the feature embedding network.
- Prepare a pre-trained model.
- Put the model under
$INSIGHTFACE_ROOT/models/. For example,$INSIGHTFACE_ROOT/models/model-r100-ii. - Run the test script
$INSIGHTFACE_ROOT/deploy/test.py.
For single cropped face image(112x112), total inference time is only 17ms on our testing server(Intel E5-2660 @ 2.00GHz, Tesla M40, LResNet34E-IR).
- TensorFlow: InsightFace_TF
- TensorFlow: tf-insightface
- TensorFlow:insightface
- PyTorch: InsightFace_Pytorch
- PyTorch: arcface-pytorch
- Caffe: arcface-caffe
- Caffe: CombinedMargin-caffe
- Tensorflow: InsightFace-tensorflow
Please check the Menpo Benchmark and Dense U-Net for more details.
Please check RetinaFace for more details.
If you find InsightFace useful in your research, please consider to cite the following related papers:
@inproceedings{deng2019retinaface,
title={RetinaFace: Single-stage Dense Face Localisation in the Wild},
author={Deng, Jiankang and Guo, Jia and Yuxiang, Zhou and Jinke Yu and Irene Kotsia and Zafeiriou, Stefanos},
booktitle={arxiv},
year={2019}
}
@inproceedings{guo2018stacked,
title={Stacked Dense U-Nets with Dual Transformers for Robust Face Alignment},
author={Guo, Jia and Deng, Jiankang and Xue, Niannan and Zafeiriou, Stefanos},
booktitle={BMVC},
year={2018}
}
@article{deng2018menpo,
title={The Menpo benchmark for multi-pose 2D and 3D facial landmark localisation and tracking},
author={Deng, Jiankang and Roussos, Anastasios and Chrysos, Grigorios and Ververas, Evangelos and Kotsia, Irene and Shen, Jie and Zafeiriou, Stefanos},
journal={IJCV},
year={2018}
}
@inproceedings{deng2018arcface,
title={ArcFace: Additive Angular Margin Loss for Deep Face Recognition},
author={Deng, Jiankang and Guo, Jia and Niannan, Xue and Zafeiriou, Stefanos},
booktitle={CVPR},
year={2019}
}
[Jia Guo](guojia[at]gmail.com)
[Jiankang Deng](jiankangdeng[at]gmail.com)