InsightFace : Additive Angular Margin Loss for Deep Face Recognition

Paper by Jiankang Deng, Jia Guo, and Stefanos Zafeiriou (Current method name ArcFace may be replaced to avoid conflicts with the company name. We will probably use the name InsightFace.)

Recent Update

2018.01.30: We provide a LResNet50E-IR model which can achieve 99.80@LFW and 97.64% at MegaFace 1M Acc. See Pretrained-Models for detail.

2018.01.29: Caffe LResNet34E-IR model is available now. We get it by converting original MXNet model to Caffe format but there's some performance drop. See Pretrained-Models for detail.

2018.01.27: MS1M clean list now available at here. Aligned facescrub images(112x112) can be downloaded here.

2018.01.26: Today we provide a pretrained LResNet34E-IR model on public drive. We also offer a simple python program to help you deploy this model to build your own face recognition application. The only requirement is using your own face detector to crop a face image before sending it to our program, no alignment needed. For single cropped face image(112x112), total inference time is only 17ms on my testing server(Intel E5-2660 @ 2.00GHz, Tesla M40, LResNet34E-IR). This model can archieve 99.65% on LFW and 96.7% on MegaFace Rank1 Acc. Please see deployment section for detail.

License

InsightFace is released under the MIT License.

Contents

0. Introduction
1. Citation
2. Requirements
3. Installation
4. How-To-Train
5. Pretrained-Models
6. Deployment
7. Results
8. Contribution
9. Contact

Introduction

  Paper link: here.

  This repository contains the entire pipeline for deep face recognition with InsightFace and other popular methods including Softmax, Triplet Loss, SphereFace and AMSoftmax/CosineFace, etc..

  InsightFace is a recently proposed face recognition method. It was initially described in an arXiv technical report. By using InsightFace and this repository, you can simply achieve LFW 99.80+ and Megaface 98%+ by a single model.

We provide a refined MS1M dataset for training here, which was already packed in MXNet binary format. It allows researcher or industrial engineer to develop a deep face recognizer quickly by only two stages: 1. Download binary dataset; 2. Run training script.

In InsightFace, we support several popular network backbones and can be set just in one parameter. Below is the list until today:

• ResNet
• MobiletNet
• InceptionResNetV2
• DPN
• DenseNet

We also support most of popular face recognition algorithms(losses), by specifying loss type:

• loss-type=0: Softmax
• loss-type=1: SphereFace
• loss-type=2: AMSoftmax/CosineFace
• loss-type=4: Ours(InsightFace)
• loss-type=12: TripletLoss

In our paper, we found there're overlap identities between facescrub dataset and Megaface distractors which greatly affects the identification performance. Sometimes more than 10 percent improvement can be achieved after removing these overlaps. This list will be made public soon in this repository.

We achieves the state-of-the-art identification performance in MegaFace Challenge, at 98%+.

Citation

If you find InsightFace useful in your research, please consider to cite our paper.

@misc{insightface2018,
  author =       {Jiankang Deng, Jia Guo and Stefanos Zafeiriou},
  title =        {Additive Angular Margin Loss for Deep Face Recognition},
  journal =      {arXiv preprint arXiv:1801.07698},
  year =         {2018}
}

If you want to download the refined MS1M dataset we provided, please cite the paper below:

@INPROCEEDINGS { guo2016msceleb,
            author = {Guo, Yandong and Zhang, Lei and Hu, Yuxiao and He, Xiaodong and Gao, Jianfeng},
            title = {M{S}-{C}eleb-1{M}: A Dataset and Benchmark for Large Scale Face Recognition},
            booktitle = {European Conference on Computer Vision},
            year = {2016},
            organization={Springer}}

If you want to download the packed VGG2 dataset we provided, please check its license here and also cite the paper below:

@article{DBLP:journals/corr/abs-1710-08092,
  author    = {Qiong Cao and
               Li Shen and
               Weidi Xie and
               Omkar M. Parkhi and
               Andrew Zisserman},
  title     = {VGGFace2: {A} dataset for recognising faces across pose and age},
  journal   = {CoRR},
  volume    = {abs/1710.08092},
  year      = {2017},
  url       = {http://arxiv.org/abs/1710.08092},
  archivePrefix = {arXiv},
  eprint    = {1710.08092},
  timestamp = {Thu, 02 Nov 2017 14:25:36 +0100},
  biburl    = {http://dblp.org/rec/bib/journals/corr/abs-1710-08092},
  bibsource = {dblp computer science bibliography, http://dblp.org}
}

Requirements

     

1. Install MXNet with GPU support(Python 2.7).

2. If you want to align dataset by yourself, install tensorflow as we're using the tf-version MTCNN. (Note that any public available MTCNN can be used to align the faces and then transform to 112x112 crop, performance/result should not change.)

Installation

1. Install MXNet by

   pip install mxnet-cu80
   

2. Clone the InsightFace repository. We'll call the directory that you cloned InsightFace as INSIGHTFACE_ROOT.

   git clone --recursive https://github.com/deepinsight/insightface.git

How-To-Train

After successfully completing the installation, you are ready to run all the following experiments.

Part 1: Dataset Downloading.

Note: In this part, we assume you are in the directory $INSIGHTFACE_ROOT/   1. Download the training set (MS1M) from the link below and place them in datasets/. Each training dataset includes following 7 files:

      faces_ms1m_112x112/
         train.idx
         train.rec
         property
         lfw.bin
         cfp_ff.bin
         cfp_fp.bin
         agedb_30.bin

The first three files are the dataset itself while the last four ones are binary verification sets.

Available training dataset(all face images are aligned and cropped to 112x112):

• Refined-MS1M@BaiduDrive, Refined-MS1M@GoogleDrive
• VGGFace2@BaiduDrive
• Any third-party contribution are always welcome, please check src/data/face2rec2.py on how to build a binary face dataset.

Part 2: Train

Note: In this part, we assume you are in the directory $INSIGHTFACE_ROOT/src/. Before start any training procedure, make sure you set the correct env params for MXNet to ensure the performance.

export MXNET_CPU_WORKER_NTHREADS=24
export MXNET_ENGINE_TYPE=ThreadedEnginePerDevice

Now we give some examples below. Our experiments were all done on Tesla P40 GPU.

1. Train our method with LResNet100E-IR.

   CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train_softmax.py --network r100 --loss-type 4 --margin-m 0.5 --data-dir ../datasets/faces_ms1m_112x112  --prefix ../model-r100

   It will output verification results of LFW, CFP-FF, CFP-FP and AgeDB-30 every 2000 batches. You can check all command line options in train_softmax.py.

   This model can achieve LFW 99.80+ and MegaFace 98.0%+

2. Train AMSoftmax/CosineFace with LResNet50E-IR.

   CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train_softmax.py --network r50 --loss-type 2 --margin-m 0.35 --data-dir ../datasets/faces_ms1m_112x112 --prefix ../model-r50-amsoftmax

3. Train Softmax with LMobileNetE.

   CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train_softmax.py --network m1 --loss-type 0 --data-dir ../datasets/faces_ms1m_112x112 --prefix ../model-m1-softmax

4. Re-Train with Triplet on above Softmax model.

   CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train_softmax.py --network m1 --loss-type 12 --lr 0.005 --mom 0.0 --per-batch-size 150 --data-dir ../datasets/faces_ms1m_112x112 --pretrained ../model-m1-softmax,50 --prefix ../model-m1-triplet

5. Train Softmax with LDPN107E on VGGFace2 dataset.

   CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7' python -u train_softmax.py --network p107 --loss-type 0 --per-batch-size 64 --data-dir ../datasets/faces_vgg_112x112 --prefix ../model-p107-softmax

Part 3: MegaFace Test

Note: In this part, we assume you are in the directory $INSIGHTFACE_ROOT/src/megaface/

1. Align all face images of facescrub dataset and megaface distractors. Please check the alignment scripts under $INSIGHTFACE_ROOT/src/align/. (We may plan to release these data soon, not sure.)

2. Next, generate feature files for both facescrub and megaface images.

   python -u gen_megaface.py

3. Remove Megaface noises which generates new feature files.

   python -u remove_noises.py

4. Start to run megaface development kit to produce final result.

Pretrained-Models

1. LResNet50E-IR@BaiduDrive, @GoogleDrive

Performance:

Method	LFW(%)	CFP-FF(%)	CFP-FP(%)	AgeDB-30(%)	MegaFace1M(%)
Ours	99.80	99.83	92.74	97.76	97.64

You can use $INSIGHTFACE/src/eval/verification.py to test all validation accuracy by pretrained models.

  2. LResNet34E-IR@BaiduDrive

Performance:

Method	LFW(%)	CFP-FF(%)	CFP-FP(%)	AgeDB-30(%)	MegaFace1M(%)
Ours	99.65	99.77	92.12	97.70	96.70

3. Caffe LResNet34E-IR@BaiduDrive, got by converting above MXNet model.

Performance:

Method	LFW(%)	CFP-FF(%)	CFP-FP(%)	AgeDB-30(%)	MegaFace1M(%)
Ours	99.46	99.60	87.75	96.00	93.29

Deployment

Note: In this part, we assume you are in the directory $INSIGHTFACE_ROOT/deploy/.

1. Download any pretrain-model above.(Or train models by yourself).

2. Put the model under $INSIGHTFACE_ROOT/models/. For example $INSIGHTFACE_ROOT/models/model-r34-amf/.

3. Check the testing script $INSIGHTFACE_ROOT/deploy/test.py then you'll know how to use it.

   Note that we do not require the input face image to be aligned but it should be cropped. We use (RNet+)ONet of MTCNN to further align the image before sending it to recognition network.

   For single cropped face image(112x112), total inference time is only 17ms on my testing server(Intel E5-2660 @ 2.00GHz, Tesla M40, LResNet34E-IR).

Results

We report the performance of LResNet100E-IR network trained on MS1M dataset with our method below:

Method	LFW(%)	CFP-FF(%)	CFP-FP(%)	AgeDB-30(%)	MegaFace1M(%)
Ours	99.80+	99.85+	94.0+	97.90+	98.0+

We report the performance of LResNet50E-IR network trained on VGGFace2 dataset with our method below:

Method	LFW(%)	CFP-FF(%)	CFP-FP(%)	AgeDB-30(%)	MegaFace1M(%)
Ours	99.7+	99.6+	97.1+	95.7+	-

Contribution

• Any type of PR or third-party contribution are welcome.

Contact

  [Jia Guo](guojia[at]gmail.com) and [Jiankang Deng](https://ibug.doc.ic.ac.uk/people/jdeng)

  Questions can also be left as issues in the repository. We will be happy to answer them.