PolyFace

PolyFace: Face Recogntion with Pytorch.

Training Data:

Our Model was trained on the cleaned CASIA Dataset (around 450K).

Preprocessing:

MTCNN face detection is used, and then a simple face alignment is applied with the 5 keypoints created by MTCNN.

Model:

A slightly modified google inception_resnet_v1 is used with input size 160*160.

Method:

Will be described in upcoming document.

Performance:

Three training logs are provided.
Currently, the best performance achieved was 99.583% on LFW data set.

Speed:

With a single Titan X Pascal, it tasks around 36 hours to finish 150,000 steps with batch size 128. With a old Titan X, training 150,000 steps takes around 47 hours.

Further information:

We are still preparing our technical document. We will release our code with the technical report.

Useful Links:

FaceNet: Face recognition using Tensorflow.
https://github.com/davidsandberg/facenet

SphereFace: Deep Hypersphere Embedding for Face Recognition, CVPR 2017.
https://github.com/wy1iu/sphereface

Contact

Xianbiao Qi: qixianbiao@gmail.com

qixianbiao/PolyFace