Real Time Face Recognition Detector. Over 30FPS on CPU!

This project is using Fast-MTCNN for face detection and TVM inference model for face recognition. At the face detection stage, the the module will output the x,y,w,h coordinations as well as 5 facial landmarks for further alignment. At the face recognition stage, the 112x112 image crop by the first stage output will be the second stage input. The output will be an 1x128 feature vector for cosine similarity measuring. The recognition pipeline can run 50FPS on CPU (2.8 GHz Quad-Core Intel Core i7).



  • OpenCV >= 3.4.1
  • TVM

Set up:

  • OpenCV
brew install opencv
brew link opencv
brew install pkg-config
pkg-config --cflags --libs /usr/local/Cellar/opencv/<version_number>/lib/pkgconfig/opencv.pc
git clone --recursive https://github.com/dmlc/tvm

mkdir build
cp cmake/config.cmake build
cd build
cmake ..
make -j4
  • tvm_complier

    Now you are able create your own .so file by using pretrained MXNet models on your own environment. Here I am using mobilefacenet-arcface model as face recognition backbone.

  • CMakeList.txt

    modify the TVM path into your own

  • Prefix: set the prefix model path to your own.

  • Recording ground truth:mkdir img and set record to 1 to record ground truth image for face recognition.


Run the project may activate your camera to capture images.

mkdir build
cd build
cmake ..
make -j4

Tun-able Parameters´╝Ü

  1. minSize: set the minimum size of faces for MTCNN detector. Larger size can ensure quick inference time.
  2. factor: set the step factor for pyramid of image. Larger factor will get fewer images after doing pyramid.
  3. Frame size: set the camera or streaming capturing frame size.
  4. Stage: set how many stage for MTCNN to implement.
  5. Average Faces: default 1


  • [ ] Quantize ArcFace
  • [ ] Quantize RetinaFace


title={Fast-MTCNN https://github.com/imistyrain/MTCNN/tree/master/Fast-MTCNN},
author={Jack Yu},

title={RetinaFace-TVM https://github.com/Howave/RetinaFace-TVM},

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},

  title={Stacked Dense U-Nets with Dual Transformers for Robust Face Alignment},
  author={Guo, Jia and Deng, Jiankang and Xue, Niannan and Zafeiriou, Stefanos},

  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},

title={ArcFace: Additive Angular Margin Loss for Deep Face Recognition},
author={Deng, Jiankang and Guo, Jia and Niannan, Xue and Zafeiriou, Stefanos},