This work addresses the problem induced by automatic gain control (AGC) when attempting to use an eigenspace method of object recognition. These methods are often re- ferred to as eigentracking, eigenmatching, eigenfaces and the like. The methods work very well when the images used as training images are all of the same exposure, and, the im- ages which are to be compared against the training images are also of the same exposure as those of the training im- ages.
Unfortunately, most digital cameras, video cameras and the like do not behave in this manner. Rather, digital imaging systems commonly apply AGC to maximize the dy- namic range of an image to make the image more appealing, and improve the images entropy. This leads to two different images being of very different exposure, defeating the effec- tiveness of the eigenspace technique because of the general compressive and non-linear nature of camera response func- tions.
Other research has proposed more robust variations
on the eigenspace technique to help with this situation, how-
ever, one may view such a practice as treating a symptom
and not the problem. Lightspace Eigentracking deals with the problem
directly. Rather than changing the error approximation in
the eigenspace technique, the problem of AGC is reduced to
that of a scalar multiplier acting on the digital image. In an
algebraic technique such as eigentracking where a singular
value decomposition is used, this scalar multiplier is easily
accommodated by the linearity of the method. Therefore,
even with the presence of AGC, the eigenspace method once
again becomes as effective as if the ideal case were present
and all images are of the same exposure.