ArcFace-MXNet-Gluon

ArcFace with MXNet Gluon

Requirements

OpenCV 4.1
scikit-image 1.15
Dlib 19.17
Tensorflow 1.14

MXNet 1.5

Installation

MMOD Face Detection

mkdir -p shared/MMOD                    
# Download the dlib face detection model
wget http://dlib.net/files/mmod_human_face_detector.dat.bz2 -P shared/MMOD
# Unpack it
bzip2 -d shared/MMOD/mmod_human_face_detector.dat.bz2

PRNet Landmark Detection

Download the pre-trained model shared by the PRNet authors
at GoogleDrive
Copy it to shared/PRNet/net-data

Data Preparation

Training Data

Deep Glint Dataset

http://trillionpairs.deepglint.com/data

The Deep Glint dataset is a combination of MS-Celeb-1M-v1c and Asian-Celeb datasets.
It contains around 6.75M images of 180,855 subjects.

|# Images |# Detected Faces|# IDs |
|:———-:|:———————:|:——-:|
|6,753,545| 6,749,639 |180,855|
Download two files train_msra.tar.gz (125GB) and train_celebrity.tar.gz (91GB)
and unpack both of them to a directory, e.g., /mnt/Datasets/Glint/images

Pre-processing

# Create a list of images
python data/dir2lst.py -i /mnt/Datasets/Glint/images -o /mnt/Datasets/Glint/glint.lst
# Run MMOD face detection
python MMOD/face_detector.py -p 8 -d /mnt/Datasets/Glint/images -i /mnt/Datasets/Glint/glint.lst
# Run PRNet landmark detection
python PRNet/landmark_detector.py -d /mnt/Datasets/Glint/images -i /mnt/Datasets/Glint/glint_dlib.lst
# Run face2rec to create a record file for mxnet
python data/face2rec.py /mnt/Datasets/Glint/glint_dlib_prnet /mnt/Datasets/Glint/images --pack-label --num-thread 8

Visualize the record

python data/show_record.py --shuffle -i /mnt/Datasets/Glint/glint_dlib_prnet.rec

Testing Data

LFW

http://vis-www.cs.umass.edu/lfw

LFW is the most popular face dataset for uncontrolled face verification,
which contains 13,233 face images of 5,748 subjects collected from the Internet.
LFW was published in 2007 and has been a standard benchmark in face recognition for a decade.
The LFW verification protocol provides 6,000 pairs of images split into ten folds.

|# Images|# Detected Faces|# IDs|
|:———:|:———————:|:—-:|
| 13,233 | 13,232 |5,748|
Download the lfw.tgz
and pairs.txt files
and unpack it to a directory, e.g., /mnt/Datasets/lfw/images

Pre-processing

# Create a list of images
python data/dir2lst.py -i /mnt/Datasets/lfw/images -o /mnt/Datasets/lfw/lfw.lst
# Run MMOD face detection
python MMOD/face_detector.py -p 8 -d /mnt/Datasets/lfw/images -i /mnt/Datasets/lfw/lfw.lst
# Run PRNet landmark detection
python PRNet/landmark_detector.py -d /mnt/Datasets/lfw/images -i /mnt/Datasets/lfw/lfw_dlib.lst
# Run face2rec to create a record file for mxnet
python data/face2rec.py /mnt/Datasets/lfw/lfw_dlib_prnet /mnt/Datasets/lfw/images --pack-label --num-thread 8
# Map the original pairs.txt to new indexes in the list file
python data/lfw2indices.py -l lfw_dlib_prnet.lst -i /mnt/Datasets/lfw

Visualize the record

python data/show_record.py --shuffle -i /mnt/Datasets/lfw/lfw_dlib_prnet.rec

Training

Training ArcFace with Glint dataset using 8 GPUs

python ArcFace/train.py --gpus 0,1,2,3,4,5,6,7 --num_workers 28 --prefix arcface-glint-nocolor --num_classes 180855
 --train_rec /mnt/Datasets/Glint/glint_dlib_prnet.rec
 --test_rec /mnt/Datasets/lfw/lfw_dlib_prnet.rec

Export trained model (optional)

python ArcFace/export.py 
 --model shared/ArcFace/arcface-glint-nocolor-best-337551.params
 --output shared/ArcFace/arcface-glint-nocolor

Testing

Testing ArcFace with LFW dataset

python ArcFace/test.py --prefix arcface-glint-nocolor
 --model shared/ArcFace/arcface-glint-nocolor-symbol.json
 --test_rec /mnt/Datasets/lfw/lfw_dlib_prnet.rec

Pretrained Models

Arcface-Glint

| Test set | Accuracy |
|:————:|:————:|
| LFW | 99.85 ± 0.20 |

Discussion