thierrypin / face-biometry

Every step for face biometry

Face Biometry

Complete pipeline for face biometry

This project contains no novelty in terms of computer vision techniques, but it integrates all the face biometry steps.

• Face detection and alignment is performed by MTCNN, with a choice of implementation from Tensorflow or MXNet. It runs in real time on both CPU and GPU.

• Feature extraction is performed by ArcFace over Mobilenet on MXNet. The vector length is 128.

Other github projects made this possible:

• From Facenet, commit 51fc8cb7880f07c766dc1cc46a6f4f619dc5626c, I got the Python/Tensorflow implementation of MTCNN.

• From Insightface, commit be3f7b3e0e635b56d903d845640b048247c41c90, I took the feature extraction model. For a compromise between speed and accuracy, I employ ArcFace + Mobilenet.

• From MXNet MTCNN Face Detection, commit b56065418b63a971fcf4f8f35d058513b0ce6cbf, I took the Python/MXNet implementation of MTCNN.

Requirements

This code requires the following packages:

• mxnet
• numpy
• opencv-python
• scikit-image
• scikit-learn
• tensorflow (optional)

Go to https://mxnet.apache.org/versions/master/install/index.html to find the adequate MXNet installation.

Tensorflow is optional, since detection defaults to MXNet. In my setup, I find that TF's implementation is a bit faster. It makes a difference if you don't have so much hardware available.

Running the demo

Create folder facerec/data/rec/ and insert face pictures to register people. The images should contain exactly one face and their names should be in the format below.

[person name]_[image number].[png|jpg|...].

Then run the demo script. Use -h to change parameters.

python3 demo.py

The base classes

For face detection, use the MTCNNDetector class, and for feature extraction use the Arcface class. For the complete pipeline, including detection, encoding and classification, use the FaceRecognition interface.