This project involved me morphing my face with a man named george. By doing this a sufficiently number of times, I was able to create a gif of a gradual morph. Later, I compute the mean neutral and mean smiling face from the FEI face database. From those means, I extrapolate a caricature of myself smiling. At the very end, I did miscellaneous things such as remorphing the caricature, changing the gender of my face, and computing the PCA of the neutral faces.
First, I had to pick the points that would define the features of the human face. I also aligned the images so that the eyes would be roughly in the same position. Experimentally, I displayed what the Delaunay triangulation over the faces would appear.
In order to compute the midway face, I first computed the midway shape by averaging the points selected on my face and george's and then calculated the Delaunay triangulation over the new points. I apply affine transformations between my and george's Delaunay triangulations to the midway's and average the colors.
I simply applied what I did to compute the mean face except I adjusted the weight of the first and second image when averaging the points and colors. One thing I did change was the process of computation in terms of code, changing from iteration to array manipulation as it was faster. Previously, it took a minute or so for a single computation. Now, it was faster to compute 45 morphs than the single morph from before. Through this, I am able to create a gif of george's face morphing into mine.
The FEI Face database had both smiling and neutral faces that already had points chosen. Below is an example of one such pairing.
I computed the mean face by first computing its shape via an average of the given data points I then morphed a few of the faces into the average shape.
By morphing all the images to the average shape and averaging the colors, I was able to compute the mean face of both neutral and smiling faces.
I warped my own face to the average shapes to get this ugly result.
Similarly, I warped the average faces to my own face shape.
I believe the warp was very distorted because my face shape is pretty different from the average shape. This is especially apparent when it comes to the warps between smiling shapes and my shape. For instance, my lips are smaller than the average face, so when it got warped, it became much larger. This was even more exacerbated by the smiling face, since the edges of the mouths are much bigger for the smile. When the smiling average was warped to my face, the mouth becomes comically small for the same reason.
I can artificially make my face smile by taking the difference between the smiling and neutral averages, then adding to my own face. This was done while in the average face shape so that they are added in the same space.
Experimentally, I warped this result back to my own face shape. I also tried first warping the delta between the smiling and neutral face first, then adding to my own face. Both yielded the same result.
Using the same technique, I tried to create a more feminized version of me and a male version of me. To get better results, I used the average chinese male and female face in order to not run into the problem I had with the brazilian dataset, that my face was too different.
First, I computed the average shape using points I selected on the average chinese male and female faces. Then, I warped all the faces to that middle shape.
Interestingly, the average male and female's faces did not change too much, whereas mine already looks a little more male. I do the same thing as I did for the previous caricature, which is adding the dx of male to female to myself in the same face shape. Then I warp it to the male face shape.
I was rather pleased with the result. I repeat the same process to create a more feminine me. This time I am less successful.
I believe the reason why it did not feminize me more is because my face is already fairly similar to the average female face. The only noticeable difference is that my eyes are slightly bigger now, but my face is also wider.
Something else I decided to try was to make an eigenbasis for faces based off of the FEI neutral faces. I simply did an SVD decomposition. Here are the first few eigenfaces.
Lastly, one thing I did was collaborate with other students to create a video of our faces all morphing into each others. This was my contribution.
