Face Morphing

Part 1: Defining Correspondences

The first step in order to generate a face morph was to define a set of Correspondences to define the matching points in my face and LeBron's face. I used the provided correspondence tool to pick the same points in both my face and LeBron's face. After doing this, I averaged these 2 points in order to get a midway shape. I then used the scikit Dealuny algorithm in order to create a triangulation of these points. I then applied this average triangulation onto the 2 faces.

Part 2: Compute the Midway Face

Now that we have a mapping from my face onto Lebron's face, I can begin working on a midway face to start the morph sequence. I did the following to calculate the average face shape. I computed the average shape by doing avg_shape = 1/2(my_face_pts + lebron_face_pts). I then calculated the inverse of the affine transformation from the triangles in the original image and the triangles in the average face. I then used the inverse morphing algorithm we learned in class in order to calculate the components of the midway image from both my face and LeBron's face. I then cross dissolved these 2 images by averaging their pixel values.

Part 3: Morph Sequence

This step was similar to the previous step. I adjusted the weights to range from 0 to 1 increasing by 0.04 per frame. This led to the following 25 frame face morph.

Part 4: Average of a Population

For this part I used the danes dataset in order to find the average of a population. This population has 37 images with correspondences between the points. I first parsed the .asf files in order to get the correspondences for each of the faces. I then took the average of all of the points. I then used the morphing code I used in the previous parts in order to map each individual onto this average danish face. Below are a few of my results.

I then averaged all of the warped images in order to get the following average image.

I then used the correspondence mapping described in the paper in order to get a correspondence of my face that maps to the one used for the danish images. I then used this correspondence in order to get a morph of my face onto the average dane's face shape and a morph of the average dane's face onto my face shape.

Part 5: Caricatures

I then used an extrapolation in order to get a face that turns me into a more than average dane or a face that turns me to less of a dane. I changed the alpha factor to -0.5 that made a picture that was a less danish version of myself or increasing it to 1.5 to get a more danish version of my face mapped onto the danish average.

Part 6: Bells + Whistles

Bells and Whistles 1: Automatic Correspondences

One of the things that I did for bells and whistles was an automatic morphing. I used the mediapipe face mesh library in order to generate correspondences automatically. Below you can see the automatic correspondes generated for my face and lebron's face.

Bells and Whistles 2: Ethnicity Morph

I used a picture of an average greek male that I found on the interenet in order to morph my face onto it. I used the midway image code that I used to create the morph between me and Lebron along with the automatic correspondence creator that I added in the previous bells and whistles part. I have attached the 3 images below.

Here is the morph of just my face shape and appearence