Practical Convolutional Neural Networks

Cascading classifiers

Even if we can extract Haar features from a particular region very quickly, it does not solve the problem of extracting such features from a lot of different places in the image; this is where the concept of cascading features comes in to help. It was observed that only 1 in 10,000 sub-regions turns positive for faces in classification, but we have to extract all features and run the whole classifier across all regions. Further, it was observed that by using just a few of the features (two in the first layer of the cascade), the classifier could eliminate a very high proportion of the regions (50% in the first region of the cascade). Also, if the sample consists of just these reduced region samples, then only slightly ...