Part of the problem with eigenfaces is that they are really sensitive to variations in pose, which, as the author notes, are present in this data set. The article doesn't actually compare the variability among these faces to the variability among a control population of Korean faces, so I don't know to interpret the results anyway, but my guess just by looking at the eigenfaces is that they are capturing pose variation in addition to variation in facial structure.
Check out the results page on the Labeled Faces in the Wild site for some better algorithms that are more robust to changes in pose, along with their performance on a sample data set: http://vis-www.cs.umass.edu/lfw/results.html. Eigenfaces are the worst algorithm tested. Amazon Mechanical Turk is better than all of the algorithms, even if you show only the face itself OR show only the remainder of the photograph that doesn't include the face.
Check out the results page on the Labeled Faces in the Wild site for some better algorithms that are more robust to changes in pose, along with their performance on a sample data set: http://vis-www.cs.umass.edu/lfw/results.html. Eigenfaces are the worst algorithm tested. Amazon Mechanical Turk is better than all of the algorithms, even if you show only the face itself OR show only the remainder of the photograph that doesn't include the face.