This section show some of my results for this project. See the methods and procedures section for more explanation of the images below.

Procrustes fit

After I got the face images and created data for each face, I had to implement the Procrustes metric. This graph shows the distance metric as a function of rotation. The fact that the graph has a minimum at about 356 degrees means that the second face should be rotated 4 degrees counterclockwise to get the best overlap by the procrustes criterion.

Average feature faces

Next I got the following average faces. I created GNUPlot PostScript pictures of the average female and male faces, plus a graph with both.

Warping one face to another face

Once we have feature data, we can take one image's feature landmarks and warp them to the location of a different image. For example, we can warp one woman's image and landmarks into my landmark locations.

Warping to gender average and overall average

As explained in the methods/procedures, once we have an average face, we can warp to it just like any other face. For each of the following rows of images, we have 1) the input face 2) the input face warped to the gender average 3) the face warped to the overall average.

Average image faces

Finally, once we've warped images to various feature point averages, we can average the warped image data to create people that look real but don't exist. The image on the left is the average female face. The image on the right is the average male face. The image in the middle is the overall average face.