This was my 1st project as part of the Udacity Computer Vision Nanodegree Program. In this project, I combined my knowledge of computer vision techniques and deep learning architectures to build a facial keypoint detection system that takes in any image with faces, and predicts the location of 68 distinguishing keypoints on each face.
From models.py, I defined the network architecture using the original Facial Keypoint Paper written by Naimish Agarwal et al. Check paper here https://arxiv.org/pdf/1710.00977.pdf
In the 1st notebook, I instantiated the CNN architecture defined in models.py. The network takes in a grayscale, square image. I applied data transform to turn input image into a normalized, square, grayscale image and then a Tensor for the model to take it as input. I selected a loss function and optimizer for training the model. The model was trained from scratch using Google Colab's compute resources.
In the 2nd notebook, I used Haar Cascade detector to detect faces in an image. Pre-processed those face images so that they are grayscaled, normalized and transformed into Tensors to be accepted as input to the trained model from Noteboook 1. The result are detected faces with facial keypoints!
Code was developed in Pytorch.