We specifically look at the transformation of these nine faces. The entire data set of 261 images was used for processing.
Special thanks to Nicole Chan, Emma Feeney, Amata Lee, Huang Xiaolei, Tim Miao, Chen Yiling for taking the yearbook photos and letting me use them.
This is the mean face of NYU Shanghai sophomores after combining 261 images.
Notice that you can faintly see glasses.
Eigenfaces depict the most important facial features of this group of faces.* We can again see that many of them feature the outline of glasses.
*They are actually the eigenvectors of the covariance matrix of the probability distribution over the vector space containing these faces (as grayscale pixel vectors).
Eigenfaces are interesting because you can approximate almost any face using them.* We show the process of reconstructing each of the original 9 faces using 1 eigenface, 2 eigenfaces, all the way to 261 eigenfaces.
*Instead of representing an image through pixels, we can represent it as a sum of eigenfaces. This is done by projecting an image vector onto the subspace spanned by a set of eigenfaces. By increasing the dimension of the subspace (using more eigenfaces, sorted by decreasing eigenvalue), this projection better approximates the original image.