03 - Face Recognition Part 1

Session 3 -Face Recognition Part 1

The goal of this assignment is to create and deploy the following models on AWS Lambda and create a static website on AWS S3 for the users to interact with the trained models:

• ResNet-34 model pre-trained on the imagenet dataset
• Custom MobileNetV2 model trained in the last assignment to classify objects belonging to a custom dataset containing following classes
  • Flying Bird
  • Large Quadcopter
  • Small Quadcopter
  • Winged Drones
• Face Alignment
• Face Swap

All the files and the models have to be deployed to AWS Lambda. The code to deploy them can be found here.

Resnet34

Pre-trained ResNet-34 trained on the ImageNet dataset. The model can be downloaded as follows

import torch
from torchvision.models import resnet

model = resnet.resnet34(pretrained=True)
model.eval()

traced_model = torch.jit.trace(model, torch.rand(1, 3, 224, 224))
traced_model.save('resnet34.pt')

Result

Input Image	Prediction
	goldfinch, Carduelis carduelis

MobileNetV2

MobileNetV2 trained on a custom dataset using transfer learning. The complete step-by-step explaination to deploy and train this model can be found here.

Result

Input Image	Prediction
	Winged Drone

Face Alignment

This model uses dlib's shape_predictor_5_face_landmarks.dat landmark detection model to align the face of an input image. The dlib model can be downloaded and extracted with the following commads

$ wget http://dlib.net/files/shape_predictor_5_face_landmarks.dat.bz2
$ bzip2 -dk shape_predictor_5_face_landmarks.dat.bz2

To execute the model, run the script
python3 face_alignment.py -p <predictor_path> -i <image_path>

Result

Input Image	Output Image

Face Swap

This model uses dlib's shape_predictor_68_face_landmarks.dat landmark detection model to swap the face in the target input image with the face in the source input image. The dlib model can be downloaded and extracted with the following commands

$ wget http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2
$ bzip2 -dk shape_predictor_68_face_landmarks.dat.bz2

Facial landmarks of the images are detected to create a mask, where the face of the other image will be placed. Initially we perform a Simple Alpha Blending and then later Seamless Cloning is done to get the final swapped image.

Result

Source Input Image	Target Input Image	Output Image