Project 3: Face Morphing

David Yi

Defining Correspondences

Here, we define correspondence points between me and George Clooney. There are 50 correspondence points on total to map the eyebrows, eyes, nose, mouth, face, and corners of each image.
I then created a Dalauney triangulation of the mean of the correspondence points.

Me & George with Mean Triangulations

Computing the "Mid-Way" Face

To compute the "Mid-Way" face of an image, we repeat the following process for each triangle in the mean shape triangulation:

Find the affine transformation T from each source triangle to the mean shape triangle
Find all points in the original image contained in the mean shape triangle through the inverse transformation. If any pixels land between two pixels, we interpolate the value from surrounding pixels
Create each new mean-shape pixel as a linear combination of the corresponding pixels from each source image
Repeat for each triangle

The Morph Sequence

To compute the Morph Sequence face of an image, we use the same algorithm that we used to generate the "Mid-Way"
face for 45 different values of alpha between [0, 1]. alpha is a variable that controls the extent of both shape warping and cross-dissolve.

The "Mean Face" of a Population

In this part, we take a bunch of Danish faces and their pre-labeled correspondence points and average them to get a mean shape. Below are 12 Danish faces that were warped to the mean Danish face.

Average Danish Face

We find the mean Danish face by averaging all of the warped faces. A surprisingly good result!!

My face warped to the mean Danish shape

It is what it is ...

Average Danish face morphed to my shape

Honestly not too bad!!

Caricatures: Extrapolating from the Mean.

Here, I create caricatures of myself by extrapolating from the mean Danish face.
The caricature shape is found by adding a scaled (alpha > 1) difference between the average Danish face and my face. Larger alpha implies greater extrapolation.