In the section, Haar features, we stated that there are 180,000 rectangle (Haar) features associated with each image sub-window. Even though each feature can be computed very efficiently with the help of an integral image, using this complete set is prohibitively expensive. In order to circumvent this predicament, a neat little trick was applied. It is reasonable to expect that not all of the 18,000 theoretically possible features within each sub-window are equally important. What if we could sample this huge set of features that we have, and for each sub-window, select a reasonably small subset of features that can help us with our classification task? If we do that, we would have a very small number of these features combined to form an effective classifier.
This is the main idea behind the technique known as AdaBoost learning. In fact, AdaBoost goes a step further and learns multiple such classifiers-each of which learn on a subset of the features. These classifiers are...