37. Features and Classification

What is a Feature?

A feature can easily be thought of as a "summarizer" of something. So features of images, are really just nice and concise summarizers of image data. Furthermore, just as how images are really just a collection of numbers in an array, features are also just another collection of numbers in an array, although usually, they are much smaller than images.

So what does this mean? Let’s be even simpler: Forget images…consider humans! An individual person has many facets about themselves and it may be hard to describe someone in totality.

What are, however, some "compact" features that we may extract about a person? Compact, because we want these features to describe something about this person, but we want this description to be a summary of what is relevant.

For example, if we wanted to place boxers into their weight class, we may want to do feature extraction on each fighter, and we would extract a two dimensional feature: height and weight (both of which are used to determine a weight class).

Those are "features" in this sense, because they nicely ignore the irrelevant; they describe a person’s weight and height, features that are useful for placing the boxer into their proper weight class, and they also ignore things like skin color, or say, hair length, etc. So, in this sense, you can think of feature extraction as a way to extract relevant information, while also smartly ignoring the irrelevant. A good feature is very succinct.

Features are distinct and measurable pieces of information in an image. And we’ll go through examples of features and how to detect them. One of the breakthroughs in computer vision came from being able to automatically come up with features that are good. However you can also do this on your own, manually.