Face alignment with a CNN extracting HOG features and realistically coloring the lips

About

This project defines, justifies and shows the results of a Convolutional Neural Network for face alignment, the task of outlining facial landmarks from the image of a face.

There is a PDF report here that discusses the solution defined.

There is a Python Notebook here that displays the data manipulation performed and the training of the CNN.

Face alignment task

1. Face alignment training and prediction methods

The flowchart of going from our training data to evaluating our face alignment predictions model.

Flowchart to get our model’s predicted points on any image (of a face).

2. Our CNN's architecture

A diagram of our CNN’s layers.

3. Results of face alignment

Plotting the 12 best and worst predictions made by our HOG-CNN model on our evaluation set with their Euclidean distance.

Lip coloring task

1. Lip coloring method

A flowchart of the method we used to colour a face's lips using the predicted points from our HOG-CNN model and colour-mapping.

2. Results of lip coloring

The best qualitative results from our lip colouring method using the HOG-CNN model’s predictions to find the lips.

The worst qualitative results from our lip colouring method using the HOG-CNN’s predictions.