ArcFace face recognition

Implementation of the ArcFace face recognition algorithm. It includes a pre-trained model based on ResNet50.

The code is based on peteryuX's implementation. Instead of using full Tensorflow for the inference, the model has been converted to a Tensorflow lite model using tf.lite.TFLiteConverter which increased the speed of the inference by a factor of ~2.27.

Installation

You can install the package through pip:

pip install arcface

Quick start

The following example illustrates the ease of use of this package:

>>> from arcface import ArcFace
>>> face_rec = ArcFace.ArcFace()
>>> emb1 = face_rec.calc_emb("~/Downloads/test.jpg")
>>> print(emb1)
array([-1.70827676e-02, -2.69084200e-02, -5.85994311e-02,  3.33652040e-03,
        9.58345132e-04,  1.21807214e-02, -6.81217164e-02, -1.33364811e-03,
       -2.12905575e-02,  1.67165045e-02,  3.52908894e-02, -5.26051633e-02,
	   ...
       -2.11241804e-02,  2.22553015e-02, -5.71946353e-02, -2.33468022e-02],
      dtype=float32)
>>> emb2 = face_rec.calc_emb("~/Downloads/test2.jpg")
>>> face_rec.get_distance_embeddings(emb1, emb2)
0.78542

You can feed the calc_emb function either a single image or an array of images. Furthermore, you can supply the image as (absolute or relative) path, or an cv2-image. To make it more clear, hear are the four possibilities:

1. (Absolute or relative) path to a single image: face_rec.calc_emb("test.jpg")
2. Array of images: face_rec.calc_emb(["test1.jpg", "test2.png"])
3. Single cv2-image: face_rec.calc_emb(cv2.imread("test.png"))
4. Array of cv2-images: face_rec.calc_emb([cv2.imread("test1.jpg"), cv2.imread("test2.png")])

The face recognition tool returns (an array of) 512-d embedding(s) as a numpy array.

Notice! This package does neither perform face detection nor face alignment! It assumes that the images are already pre-processsed!

Benchmark

Model	Backbone	Framework	LFW Accuracy	Speed [ms/embedding] *
ArcFace paper	R100	MXNet	99.82	-
ArcFace TF2	R50	Tensorflow 2	99.35	102
This repository	R50	Tensorflow Lite	96.87	45

* executed on a CPU: Intel i7-10510U

License

Licensed under the EUPL, Version 1.2 or – as soon they will be approved by the European Commission - subsequent versions of the EUPL (the "Licence"). You may not use this work except in compliance with the Licence.

License: European Union Public License v1.2

Acknowledgement

This work has been carried out within the scope of Digidow, the Christian Doppler Laboratory for Private Digital Authentication in the Physical World, funded by the Christian Doppler Forschungsgesellschaft, 3 Banken IT GmbH, Kepler Universitätsklinikum GmbH, NXP Semiconductors Austria GmbH, and Österreichische Staatsdruckerei GmbH and has partially been supported by the LIT Secure and Correct Systems Lab funded by the State of Upper Austria.