[Paper Review] Eigenfaces for Recognition

This paper talked about the application of principal component analysis to face detection and identification. It seems that this is the first complete work in applying PCA to face recognition, although such techniques has become quite standard (and somehow old and out-of-date) now. Anyway, we should bear in mind it is a 1991 paper. In that time, even acquiring digital color videos can be awesome for the public.

The system comprises two main components: face detection and face classification. The author proposes to use PCA to extract key directions that best explaining the variation of face images in high dimensional space. Those eigenvectors obtained are supposed to be the most important factors that constitutes a face. Thus, by projecting unknown image to this reduced "face space", one can then judge if an image is a face from the projected length in each direction. Furthermore, in case that the image is indeed a face, we may compare it in face space with the pre-built database of individual faces to identify its personality. Those are the general ideas of the paper.

Admittedly, many parts of the system are obsolete from today's perspective. For example, Adaboost seems to be the most successful method in face detection now. By using integral images that accelerates Haar-feature computation and by filter cascading, it is shown in [Viola 04] that accurate face detection of various sizes can be done real-time on conventional PC. It is unlikely correlation-based methods like subspace projection can achieve the same performance. On the other hand, there were also many methods that better suit face identification proposed later (for ex., the fisher-face). The value of this paper should be in the idea to use principal components to deal with unknown variations of high dimensional data, which certainly benefits subsequent researches like Active Appearance Model. Another vital contribution of this paper, I think, is a practical method to perform eigenvector decomposition on relative small number of observations from high dimensional space. This is critical even in today.