The point sets that I used are here.
The sequence of frames in my part of the class morph (from cmcamero to gsj) is here.
I turned this morph into a movie which can be seen here. Apologies for the crap encoding.
The sequence of frames morphing me to the mean face is here. I turned this morph into a movie which can be seen here. The mean face can be seen here.
This caricature was produced by morphing my geometry away from the mean geometry. Well... at least it makes my brain look bigger.
Singular value decomposition produced this orthonormal basis of the texture space.
If the background of the images is discarded, the SVD produces this basis of texture space.
The following picture shows the expressiveness of the eigenvectors for the texture and geometry space. In the lower-left is my face, reconstructed from the full eigenspace. As images progress to the right, successively more geometry eigenvectors are removed, and my geometry is orthogonally projected onto the resulting space. As images progress upwards, successively more texture eigenvectors are removed, and my texture is orthogonally projected onto the resulting space. The image in the upper-right contains only one geometry and texture dimensions, namely that of the mean face. In each step in each dimension of this image, three eigenvectors are stripped out.
It should be noted that the background (the wall) is not an important part of the texture space for faces. The following pictures shows the same content as the above, but with the background removed. It should be noted that the transition of the textures is more smooth in this representation.
When SVD is used to create a basis for the space of faces in the class with mine excluded, and then my face is reconstructed by orthogonal projection into this space, the result is the following. On the left is the reconstructed image. On the right is the target image. The space is not as expressive as one might hope, but the approximation is not terrible.
A decomposition of this image adding sequentially more eigenvectors is shown here.
