Cascade classifiers

In this section, we will discuss the powerful cascade classifier and its components, Haar features, integral images, Adaptive Boosting (Adaboost), and cascading to build an object detector.

In a nutshell, to construct an object detector, you train it using positive samples (let's say, faces of size 24x24) and negative samples (any other images that are not faces). You keep refining the training process to minimize the training error (the total number of faces classified as non-faces and total number of non-faces classified as faces).

Once the training is done and we get a new image, we ask the detector to check if it has a positive sample (that is, face). The steps followed to do so are as follows:

1. The detector will scan the input image using a scanning window, and every window scanned will get a score.
2. The detector then will say that this window contains a positive sample if its score is greater than a certain threshold; otherwise, it does not.

Haar-like features

Haar...