Face Morphing

Daniel Geng

Midway Face

We are going to morph George and Brad’s faces, both images shot by photographer Martin Schoeller:

George Clooney

Brad Pitt

We first calculate the midway face by defining corresponding points and calculating the average shape of the two faces. We then morph both faces into the average shape and take an average. The result:

Midway Face (the eyes were a tiny bit misaligned)

Some amusing failure cases:

Morph Sequence

To produce the full morph sequence, we simply follow the same procdure as the midway face, but instead of taking an exact average we take a weighted one, with different weights resulting in different frames of the morph.

Brad to George

Brad to George (large)

The smaller morph sequence looks much better, probably because the downsampling makes it hard to see the shoddy correspondence points that I made.

“Mean” Face of a Population

We now use our warp function to warp faces from a dataset into the average shape of the dataset. Then by taking an average we obtain a “population average.”

We use the males from the Danish face dataset (which unfortunately only has points labeled on facial features, with no head shape features.

The average male face in the dataset

Here are some samples of warpped faces:


Finally, we move a picture of myself closer to the population mean. The warpped picture looks a tad freaky because I’m pretty bad at selecting points.

(Mouse over for caricatured image)

Interestingly enough rather than making me look more male the population mean just makes me look slightly more “whiter,” which makes sense given that this is a set of Danish faces.