Face Unlock

This project was done in my sophomore year at IvLabs to implement face-recognition algorithms from scratch, and further plan to deploy as face recognition based door lock. During this, we implemented Triplet Network and FaceNet algorithms with ResNet as the backbone architecture implemented from scratch and performed one-shot and few-shot learning on different datasets.

ABD

Created on December 7|Last edited on February 10

Comment

﻿

﻿
﻿
Triplet Network
﻿
﻿﻿﻿
Loss Function: 
L=∑iN[∥f(xia)−f(xip)∥22−∥f(xia)−f(xin)∥22+α]+L = \sum_i^N [\|f(x^a_i)-f(x^p_i)\|^2_2 - \|f(x^a_i)-f(x^n_i)\|^2_2 + \alpha]_+L=∑iN​[∥f(xia​)−f(xip​)∥22​−∥f(xia​)−f(xin​)∥22​+α]+​﻿
	where, xax^axa﻿ = anchor image, xpx^pxp﻿ = positive sample, xnx^nxn﻿ = negative sample
﻿
Datasets
Sweeps﻿
Sweep: vq62ojvw 115
﻿
﻿
Results﻿
Run: Last Run1
﻿
﻿
﻿
Run: Run_09-Dec 06:351
﻿

ArchitechtureEmbeddings
 DimensionNo. of Learnable 
ParametersEpochsLearning RateBatch Size
ResNet-1812811,242,1762000.002 
 (Reduced by factor of 2 every 50 epochs)256
﻿

AccuracyPrecisionRecallROC 
 Area Under CurveEuclidean 
 Distance TAR @ FAR=1e-2
88.35%88.46%88.23%0.95081.10461.07%
﻿
﻿
Project OutcomeLearned about the One-Shot learning approach on the various datasets.
Implemented Siamese Nets and FaceNet models using a ResNet backbone implemented from scratch. 
Achieved 88.35% accuracy on the Labelled Faces in the Wild dataset.
Contributed to a dataset class for LFW in the Torchvision library.
﻿
References
BibTexIf you find this work useful please cite this report.
 @misc{shaikh_2021,
	title	={Face unlock}, 
	url	={https://wandb.ai/abd1/Face-Unlock/reports/Face-Unlock--VmlldzoxMzEyNjQ4}, 
	journal	={W&amp;B Report},
	author	={Shaikh, Muhammed Abdullah}, 
	year	={2021}, month={Dec}
} 
﻿
﻿

Architechture	Embeddings Dimension	No. of Learnable Parameters	Epochs	Learning Rate	Batch Size
ResNet-18	128	11,242,176	200	0.002 (Reduced by factor of 2 every 50 epochs)	256

Accuracy	Precision	Recall	ROC Area Under Curve	Euclidean Distance	TAR @ FAR=1e-2
88.35%	88.46%	88.23%	0.9508	1.104	61.07%

Add a comment