arcface

ArcFace

input

• correct pair
  

• incorrect pair
  

(Image from ronghuaiyang/arcface-pytorch#63)

Input the original image1 and its inversion, and the original image2 and its inversion.
Invered images are generated automatically.
(All images are treated as grayscale images)

• Ailia input Shape: (4, 1, 128, 128)
• Range: [-1, 1]

output

A similarity of a pair of images.

usage

Automatically downloads the onnx and prototxt files on the first run.
It is necessary to be connected to the Internet while downloading.

By default, the following two images are loaded: correct_pair_1.jpg, correct_pair_2.jpg

$ python3 arcface.py
...
Similarity of (correct_pair_1.jpg, correct_pair_2.jpg) : 0.5981666445732117
They are the same face!

If you want to specify images, specify the paths of the two images after the --inputs option.

$ python3 arcface.py --inputs IMAGE_PATH1 IMAGE_PATH2

By adding the --video option, you can compare the face of the video and calculate the similarity. If you pass 0 as an argument to VIDEO_PATH, you can use the webcam input instead of the video.

$ python3 arcface.py --video VIDEO_PATH

$ python3 arcface.py --video 0

Reference

arcface-pytorch

Framework

PyTorch 1.3.1

Model Format

ONNX opset = 10

Netron

arcface.onnx.prototxt