Note

This report contains a gif that may not load properly if viewed in the pdf. Please click the link above to view the report online.

Overview

This project explores how to create a face morph animation by warping the two images into a "mid-way" face. I used Delaunay triangulation to find the most optimal triangles to warp the images. Finally, I then used the affine transformation to warp the images and blend them together to create caricatures and mean face of a population.

Defining Correspondences

First, I used the labelling tool provided by last year's students to define the pairs of corresponding points that mapped both images' eyes, mouths, noses, and face structure. I had also added corner points for both images to help with the warping process. After I got the points, I passed them into Delaunay to find the optimal triangles. Finally, I overlayed the points and the triangles over the original images. Here are the images with the corresponding points and triangulations:

Affine Warping

Once I had the triangles for the correspondences, I first averaged them to get the "mid-way" points. I then used the equations to inverse warp from the "mid-way" points back to the original images' points. During this process, I calculated the affine transformation matrix so that I could fill in the "mid-way" face with the points from the original images. I had also incorporated warp_frac to interpolate the location of the points between the images and dissolve_frac to interpolate the resulting colors between the two original images. Here are the original images and the "mid-way" face:

The Morph Sequence

After morphing my face with Steph Curry's, I decided to vary the warp_frac and dissolve_frac to create a morph sequence. These two parameters controlled how much warping and cross dissolving between the two images and ranged from 0 to 1. I had also created a video of the morph sequence to show the gradual change between the two faces. For the video, I created 55 equally spaced frames, used an fps of 20, and also reversed it to show Steph Curry morphing into my face as well. Here is the video:

The Mean Face

I used the Danes dataset to get the mean face of the population. This was very similar to the previous merging of two faces but instead I would take in a list of images and points and equally weight them for the final merged image. I found the average face of this population and then warped some of the original Danes to this average face. Here are the images:

I also warped my face onto this average Dane face and that face onto mine. Here are the results:

Extrapolating from the Mean

For the caricature, I used the mean neutral Dane face from the last step and extrapolated away from it. The equation I used was: $$ \alpha * \text{Ian} + (1 - \alpha) * \text{average Dane} $$ I set $\alpha$ to be larger than 1 to extrapolate in the direction of Ian's face. Here are the images:

Morphing to White Guy

For my bells and whistles, I decided to morph my face with an average white guy's face that I had found online. To get just the shape of the other white guy, I used a warp_frac of 0 and a dissolve_frac of 0.6. Then, to get just the color, I used a warp_frac of 1 and a dissolve_frac of 0.2. Finally, I used a warp_frac of 0.5 and a dissolve_frac of 0.5 to get both a mix of the shape and color. Here are the images: