Tuesday, March 4, 2008

[Reading] Eigenfaces to Recognition

I'm surprised by the content of this paper. I've heard eigenface for a long time, I didn't realize the true importance of this work. (And I find this is another Pentland's great paper)

In this paper an efficient way to compute eigenvectors is used (maybe not the first one if we consider other research fields) and many possible attacks to the system are examined. The first topic is very important because it provides us a way to perform PCA on high-dimensional data, as long as the number of the samples is much smaller than the dimension of the samples. The second topic is about the performance degradation due to scaling, lighting, and rotation (on the image domain, not the rigid motion of the head). The authors also discuss some possible ways to solve these problems. For example, using a multi-scale processing.

Another interesting thing is that no one is talking about neural networks anymore, but obviously it was hot at 1991.

2 comments:

AcmeChimera said...

Yeah, but why neural networks die out? After all it sounds to be attractive simply by its name.

Chia-Kai Liang said...

There are many dead techniques with nice names...