A Common GD Based Algorithm
(and its shortcomings)

A common Gabor Decompositon based pattern-recognition algorithm can be schematically described as follows:
 

Training

An appearance-based model of the pattern of interest (in this case, an eye) is encoded as a vector of Gabor responses.

This realistically involves averaging over a certain number of subjects (the training set).

More generally, response vectors from the images in the training set could be used to train a classifier.

Search

When looking from the pattern on a new image, a Gabor feature vector is extracted at each point.

The vector is compared against the model. The resulting distance map is expected to have a minimum at the location of the pattern we are looking for, were the feature vector best resembles the model.

Here again, the distance map could represent the output of a classifier rating each image pixel.

Disadvantages

The main problem with this approach is that the Gabor decomposition must be computed everywhere in the image. This requires, in practice, working in the Fourier domain and is therefore time consuming.

This common algorithm as it stands is therefore ill-suited for active vision applications.

Follow this link to the Saccadic Search Home Page


Back to F. Smeraldi's Home Page
Last modified Feb 11th, 1999