Learning OpenCV 3

Chapter 16. Keypoints and Descriptors

Keypoints and the Basics of Tracking

This chapter is all about informative feature points in images. We will begin by describing what are called corners and exploring their definition in the subpixel domain. We will then learn how to track such corners with optical flow. Historically, the tracking of corners evolved into the theory of keypoints, to which we will devote the remainder of this chapter, including extensive discussion of keypoint feature detectors and descriptors implemented in the OpenCV library for you to use.¹

The concept of corners, as well as that of keypoints, is based on the intuition that it would be useful in many applications to be able to represent an image or object in an invariant form that will be the same, or at least very similar, in other similar images of the same scene or object. Corner and keypoint representations are powerful methods for doing this. A corner is a small patch of an image that is rich in local information and therefore likely to be recognized in another image. A keypoint is an extension of this concept that encodes information from a small local patch of an image such that the keypoint is highly recognizable and, at least in principle, largely unique. The descriptive information about a keypoint is summarized in the form of its descriptor, which is typically much lower-dimensional than the pixel patch that formed the keypoint. The descriptor represents that patch so as to make it much easier ...