Face-Recognition

This repository consists of code to fine-tune popular face-recognition architectures with LFW and QMUL-Survface datasets for evaluationg Low Resolution Face Recognition. This project is carried out as part of my Masters Thesis - "Bias and Fairness in Low Resolution Image Recognition" under the guidance of Dr. Mayank Vatsa and Dr. Richa Singh

Requirements:

Clone the repository

git clone https://github.com/ksasi/face-recognition.git

Install using pip

pip install -r requirements.txt

Datasets:

Dataset	Description
LFW	Labeled Faces in the Wild is a public benchmark that can be used for face recognition and verification.LFW deep funneled imagesconsists of LFW images aligned using deep funneling.
QMUL-SurvFace	This consists of native low-resolution face images. These are not synthesised by artificial down-sampling of high-resolution face images.

Architectures:

Fine-tuning - Contrastive Loss
(LightCNN29, ArcFace, VGGFace2)

Architecture	Description
LightCNN29	This is a 29 layer CNN model, where a variation of maxout activation known as Max- Feature-Map (MFM) is introduced in each convolution layer. This is trained with large scale noisy face datasets such as CASIA-WebFace and MS-Celeb-1M in gray-scale.
VGGFace2	This is a resnet50 backbone trained with MS-Celeb-1M and the fine-tuned with VG- GFace2 dataset.
ArcFace	This is a resnet18 backbone pre-trained with MS1MV3 dataset with Arcface i.e. Additive angular margin loss for deep face recognition.

Inference
(LightCNN29, ArcFace, VGGFace2)

Experimental Setup:

• Download pretrained model checkpoints for LightCNN29, VGGFace2 and ArcFace
• Place the checkpoint files in the "models" folder
• Download the Processed LFW and QMUL-SurvFace datasets
• Uncompress the downloaded lfw-deepfunneled_processed.tar.gz and QMUL-SurvFace-v1.zip into the data folder
• Download and uncompress processed mated_probe.zip and processed gallery.zip into "data/QMUL-SurvFace/Face_Identification_Test_Set" location from here

Fine-tuning and Evaluation Scripts:

fine_tune.sh script in scripts folder can be executed as shown below:

Fine-tuning with LFW dataset

1. Fine-tune and Evaluate pretrained LightCNN29 model with LFW dataset

   sh fine_tune.sh LightCNN29 LFW

2. Fine-tune and Evaluate pretrained VGGFace2 model with LFW dataset

   sh fine_tune.sh VGGFace2 LFW

3. Fine-tune and Evaluate pretrained ArcFace model with LFW dataset

   sh fine_tune.sh ArcFace LFW

Fine-tuning with QMUL-SurvFace dataset

1. Fine-tune and Evaluate pretrained LightCNN29 model with QMUL-SurvFace dataset

   sh fine_tune.sh LightCNN29 SurvFace

2. Fine-tune and Evaluate pretrained VGGFace2 model with QMUL-SurvFace dataset

   sh fine_tune.sh VGGFace2 SurvFace

3. Fine-tune and Evaluate pretrained ArcFace model with QMUL-SurvFace dataset

   sh fine_tune.sh ArcFace SurvFace

Results:

Identification Accuracy:

Model	Rank1	Rank1	Rank10	Rank10
Dataset	LFW	SurvFace	LFW	SurvFace
LightCNN29	96.8	6.3	98.6	15.2
VGGFace2	95.6	1.6	99.7	5.7
ArcFace	80.7	1.8	89.6	5.8

CMC Curves:

LightCNN29 - LFW	LightCNN29 - QMUL-SurvFace

VGGFace2 - LFW	VGGFace2 - QMUL-SurvFace

ArcFace - LFW	ArcFace - QMUL-SurvFace

Contact:

For questions and clarifications, please contact @ksasi or raise an issue on GitHub.

References:

The code is adapted from the following repositories:

1. Light CNN for Deep Face Recognition, in PyTorch
2. VGGFace2 Dataset for Face Recognition
3. InsightFace: 2D and 3D Face Analysis Project
4. PyTorch Metric Learning

How to cite this repository

If you used this repository in your work, please cite it as below:

@software{face-recognition,
  author = {Sasikanth, Kotti},
  title = {{face-recognition}},
  year = {2022},
  url = {https://github.com/ksasi/face-recognition}
}