InsightFace: 2D and 3D Face Analysis Project

License

The code of InsightFace is released under the MIT License.

Recent Update

2018.12.13: TVM-Benchmark

2018.10.28: Gender-Age created with a lightweight model. About 1MB size, 10ms on single CPU core. Gender accuracy 96% on validation set and 4.1 age MAE.

2018.10.16: We got rank 1st on IQIYI_VID(IQIYI video person identification) competition which in conjunction with PRCV2018, see detail.

2018.07.17: Model-Zoo, Dataset-Zoo

2018.06.14: There's a large scale Asian training dataset provided by Glint, see this discussion for detail.

2018.05.16: A new training dataset released here which can easily achieve much better accuracy. See discussion for detail.

2018.04.23: Our implementation of MobileFaceNet is now available. Please set --network y1 to use this lightweight but powerful backbone.

2018.03.26: We can train with combined margin(loss-type=5), see Verification Results On Combined Margin.

2018.02.13: We achieved state-of-the-art performance on MegaFace-Challenge. Please check our paper and code for implementation details.

Deep Face Recognition

Introduction
Training Data
Train
Pretrained Models
Verification Results On Combined Margin
Test on MegaFace
512-D Feature Embedding
Third-party Re-implementation

Deep Face Recognition

Introduction

In this repository, we provide training data, network settings and loss designs for deep face recognition. The training data includes the normalised MS1M and VGG2 datasets, which were already packed in the MxNet binary format. The network backbones include ResNet, InceptionResNet_v2, DenseNet, DPN and MobileNet. The loss functions include Softmax, SphereFace, CosineFace, ArcFace and Triplet (Euclidean/Angular) Loss.

loss-type=0: Softmax
loss-type=1: SphereFace
loss-type=2: CosineFace
loss-type=4: ArcFace
loss-type=5: Combined Margin
loss-type=12: TripletLoss

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.

Training Data

All face images are aligned by MTCNN and cropped to 112x112:

Please check Dataset-Zoo for detail information and downloading.

Note: If you use the refined MS1M dataset and the cropped VGG2 dataset, please cite the original papers.

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.

Train

Install MXNet with GPU support (Python 2.7).

pip install mxnet-cu80

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) and place it in $INSIGHTFACE_ROOT/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 training dataset while the last four files are verification sets.

Train deep face recognition models. In this part, we assume you are in the directory $INSIGHTFACE_ROOT/src/.

export MXNET_CPU_WORKER_NTHREADS=24
export MXNET_ENGINE_TYPE=ThreadedEnginePerDevice

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_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 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). Fine-turn the above Softmax model with Triplet loss.

CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train_triplet.py --network m1 --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 LDPN107E network with Softmax loss 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

Verification results.

LResNet100E-IR network trained on MS1M dataset with ArcFace loss:

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

LResNet50E-IR network trained on VGGFace2 dataset with ArcFace loss:

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

We report the verification accuracy after removing training set overlaps to strictly follow the evaluation metric. (C) means after cleaning

Dataset	Identities	Images	Identites(C)	Images(C)	Acc	Acc(C)
LFW	85742	3850179	80995	3586128	99.83	99.81
CFP-FP	85742	3850179	83706	3736338	94.04	94.03
AgeDB-30	85742	3850179	83775	3761329	98.08	97.87

Pretrained Models

You can use $INSIGHTFACE/src/eval/verification.py to test all the pre-trained models.

Please check Model-Zoo for more pretrained models.

LResNet50E-IR@BaiduDrive, @GoogleDrive Performance:

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

LResNet34E-IR@BaiduDrive Performance:

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

Caffe LResNet50E-IR@BaiduDrive, converted by above MXNet model. Performance:

Method	LFW(%)	CFP-FF(%)	CFP-FP(%)	AgeDB-30(%)	MegaFace1M(%)
Ours	99.74	-TBD-	-TBD-	-TBD-	-TBD-

Verification Results on Combined Margin

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=0.9, m2=0.4, m3=0.15, run following command:

CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train_softmax.py --network r100 --loss-type 5 --margin-a 0.9 --margin-m 0.4 --margin-b 0.15 --data-dir ../datasets/faces_ms1m_112x112  --prefix ../model-r100

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

Test on MegaFace

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

Note: We found there are overlap identities between facescrub dataset and Megaface distractors, which significantly affects the identification performance. This list is released under $INSIGHTFACE_ROOT/src/megaface/.

Align all face images of facescrub dataset and megaface distractors. Please check the alignment scripts under $INSIGHTFACE_ROOT/src/align/.
Generate feature files for both facescrub and megaface images.

python -u gen_megaface.py

Remove Megaface noises which generates new feature files.

python -u remove_noises.py

Run megaface development kit to produce final result.

512-D Feature Embedding

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-r34-amf.
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).

Third-party Re-implementation

TensorFlow: InsightFace_TF
TensorFlow: tf-insightface
PyTorch: InsightFace_Pytorch
PyTorch: arcface-pytorch
Caffe: arcface-caffe

Face Alignment

Todo

Face Detection

Todo

Citation

If you find InsightFace useful in your research, please consider to cite the following related papers:

@article{deng2018arcface,
title={ArcFace: Additive Angular Margin Loss for Deep Face Recognition},
author={Deng, Jiankang and Guo, Jia and Zafeiriou, Stefanos},
journal={arXiv:1801.07698},
year={2018}
}

Contact

[Jia Guo](guojia[at]gmail.com)
[Jiankang Deng](jiankangdeng[at]gmail.com)

songshuaijin/insightface

InsightFace: 2D and 3D Face Analysis Project

License

Recent Update

Contents

Deep Face Recognition

Introduction

Training Data

Train

Pretrained Models

Verification Results on Combined Margin

Test on MegaFace

512-D Feature Embedding

Third-party Re-implementation

Face Alignment

Face Detection

Citation

Contact