CS180 Project 3: Face Morphing

Part 1: Triangulation

For the face morphing, I’m using a selfie and george clooney’s picture.

I started by creating one-to-one point correspondence on key features of the 2 images using the tool linked to on the project page (this one). Moreover, I added the points for the 4 corners of the images after loading in the images. After getting the points, I used Delaunay triangulation to create the triangle meshes. Here are the results:

Original Images: my selfie and mr clooney

Triangle mesh generated using Delaunay algorithm.

Part 2: Mid-way Face

To compute a mid-way face, I start by getting the inverse of the affine transformation between 2 corresponding triangles. I directly compute the INVERSE because we are doing inverse warping. To get the inverse affine transformation, I did:

T \cdot A = A'\\ T = A' \cdot A^{-1}\\ T^{-1} = (A' \cdot A^{-1})^{-1}\\ T^{-1} = (A^{-1})^{-1} \cdot (A')^{-1}\\ T^{-1} = A \cdot (A')^{-1}

$T$ is the affine transformation, $A$ is the stacked coordinates of the original triangle with the bias appended, and $A'$ is the stack coordinates of the target triangle with the bias appended. Here’s the mid-way face of me and george clooney:

Part 3: The Morph Sequence

I made 2 gifs, one with 46 frames and one with 90 frames to see if more frames will give smoother results. The morph sequences are uploaded to YouTube.

https://youtube.com/shorts/K_1adS11cvU?feature=share

https://youtube.com/shorts/Ze-0xWLgcHY?feature=share

Part 4: Population Average

I used the FEI dataset to calculate a population average. More specifically, I used the frontalimages_spatiallynormalized images and the corresponding point annotations. I computed 2 mean faces of the dataset, one using the neutral expression images and one using the smiling images. Let’s see the results:

Here are some images in the dataset with their face geometry warped into the average shape:

I tried taking a new image of myself to make it more aligned with the images of the dataset:

Here’s my face warped into the average geometry and the avg face warped into my geometry:

Definitely unexpected results. I think it can be attributed to the difference in camera angles and setup. I couldn’t really get a picture that is really similar to the FEI dataset.

Part 5: Caricature:

Made a caricature of myself using wrap_frac=-0.6:

Geez i know my forehead is huge, but no need to expose me like that lol

Here are some other caricatures using images from the dataset:

Part 6: B & W, Changing My Gender

I found an average image of Chinese female online for this part. Here’s the original image:

And the here’s the gender change results:

Future improvements might focus on the hair features. If the female hair style is more accentuated, I think the results will look even better.