April 2017
Beginner to intermediate
250 pages
6h 28m
English
Despite the ease of which discriminative patch models can be learned as described previously, it is worth considering whether generative patch models and their corresponding training regimes are simple enough to achieve similar results. The generative counterpart of the correlation patch model is the average patch. The learning objective for this model is to construct a single image patch that is as close as possible to all examples of the facial feature as measured via the least-squares criterion:

The solution to this problem is exactly the average of all the feature-centered training image patches. ...
Read now
Unlock full access