Compare face detectors - Dlib, OpenCV, Others..

We are neighborhood

Test image size : HD (720p)

We wanted to check processing time on same condition. but It couldn't becasue each method demand different input size. (ex. opencv dnn use 300x300 bgr image.)

So, Each code has a different image size.

ocv-dnn : 300x300
ocv-haar, dlib-hog, dlib-cnn, fr-hog, fr-cnn : VGA(640x360)
mtcnn : HD(1280x720)
s3fd : HD --> 1/8 scale. low resolution but awesome performance!
insightface(retianface_r50_v1) : VGA(640x360)

Test on Intel i7-6700K & GTX1080.

Test on Intel Xeon E5-1660 & NVIDIA GV100.

Test on MacBook pro 2018 i5.

• Python 3.7
• OpenCV 4.1.1 (option: build from src with highgui)
• Dlib 19.17.0
• face_recognition 1.2.3
• pytorch 1.2.0
• mxnet-cu100 1.5.1.post0

First, install libs

pip install opencv-contrib-python
pip install torch
pip install dlib
pip install face_recognition
pip install easydict
pip install mxnet-cu100
pip install insightface

Second, prepare weight file (s3fd)

download s3fd weight: https://drive.google.com/open?id=1Dyr-s3mAQEj-AXCz8YIIYt6Zl3JpjFQ7

[ROOT DIR]/S3FD/weights/s3fd.pth

Last, check run-time for each algorithm.

./run.sh

Of course, You can execute each file. and watch the result image (need opencv high gui)

python dlib-hog.py

opencv haar cascade

• https://docs.opencv.org/3.4.1/d7/d8b/tutorial_py_face_detection.html

opencv caffe based dnn (res-ssd)

• https://github.com/opencv/opencv/tree/master/samples/dnn

dlib hog

• http://dlib.net/

dlib cnn

• http://blog.dlib.net/2016/10/easily-create-high-quality-object.html

face-recognition (dlib-based)

• https://github.com/ageitgey/face_recognition

mtcnn

• https://github.com/TropComplique/mtcnn-pytorch (code)
• https://arxiv.org/abs/1604.02878 (paper)

s3fd

• https://github.com/tkwoo/S3FD.pytorch (code: editing pytorch 1.3 style)
• https://github.com/yxlijun/S3FD.pytorch (original code)

insightface(retinaface)

• https://github.com/deepinsight/insightface (code)