InsightFace

PyTorch implementation of Additive Angular Margin Loss for Deep Face Recognition. paper.

@article{deng2018arcface,
title={ArcFace: Additive Angular Margin Loss for Deep Face Recognition},
author={Deng, Jiankang and Guo, Jia and Niannan, Xue and Zafeiriou, Stefanos},
journal={arXiv:1801.07698},
year={2018}
}

Dataset

Introduction

MS-Celeb-1M dataset for training, 3,804,846 faces over 85,164 identities.

Dependencies

• Python 3.6.8
• PyTorch 1.3.0

Usage

Data wrangling

Extract images, scan them, to get bounding boxes and landmarks:

$ python extract.py
$ python pre_process.py

Image alignment:

1. Face detection(MTCNN).
2. Face alignment(similar transformation).
3. Central face selection.
4. Resize -> 112x112.

Original	Aligned & Resized	Original	Aligned & Resized

Train

$ python train.py

To visualize the training process：

$ tensorboard --logdir=runs

Performance evaluation

LFW

Introduction

Use Labeled Faces in the Wild (LFW) dataset for performance evaluation:

• 13233 faces
• 5749 identities
• 1680 identities with >=2 photo

Download

Download LFW database put it under data folder:

$ wget http://vis-www.cs.umass.edu/lfw/lfw-funneled.tgz
$ wget http://vis-www.cs.umass.edu/lfw/pairs.txt
$ wget http://vis-www.cs.umass.edu/lfw/people.txt

Start evaluation

$ python lfw_eval.py

Results

Note(*): with 1 Nvidia Tesla P100.

theta j Distribution

Error analysis

See also LFW Face Database Errata

False Positive

2 false positives:

1	2	1	2

False Negative

8 false negative:

1	2	1	2

MegaFace

Introduction

MegaFace dataset includes 1,027,060 faces, 690,572 identities. Link

Challenge 1 is taken to test our model with 1 million distractors.

Download

1. Download MegaFace and FaceScrub Images
2. Download Linux DevKit from MagaFace WebSite then extract to megaface folder:

$ tar -vxf linux-devkit.tar.gz

Generate features

1. Crop MegaFace.
2. Generate features for FaceScrub and MegaFace.
3. Remove noises. Note: we used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface.

$ python3 megaface.py --action crop_megaface

$ find megaface/facescrub_images -name "*.bin" -type f -delete
$ find megaface/MegaFace_aligned/FlickrFinal2 -name "*.bin" -type f -delete

$ python3 megaface.py --action gen_features

Evaluation

Start MegaFace evaluation through devkit:

$ cd megaface/devkit/experiments
$ python run_experiment.py -p /dev/code/mnt/InsightFace-v2/megaface/devkit/templatelists/facescrub_uncropped_features_list.json /dev/code/mnt/InsightFace-v2/megaface/MegaFace_aligned/FlickrFinal2 /dev/code/mnt/InsightFace-v2/megaface/facescrub_images _0.bin results -s 1000000

Results

Curves

Draw curves with matlab script @ megaface/draw_curve.m.

CMC	ROC

Textual results

Done matching! Score matrix size: 3379 972313
Saving to results/otherFiles/facescrub_megaface_0_1000000_1.bin
Computing test results with 1000000 images for set 1
Loaded 3379 probes spanning 80 classes
Loading from results/otherFiles/facescrub_facescrub_0.bin
Probe score matrix size: 3379 3379
distractor score matrix size: 3379 972313
Done loading. Time to compute some stats!
Finding top distractors!
Done sorting distractor scores
Making gallery!
Done Making Gallery!
Allocating ranks (972393)

Rank 1: 0.964733