Characteristic Points

The identification (extraction) of keypoints (keypoint detection), their characterization (feature description), and finally comparison (matching) are closely related topics within the field of computer vision. Applications that utilize keypoints range from panoramic image creation to three-dimensional reconstruction, from visual odometry to object tracking, and in many other use cases.

The concept of a keypoint emphasizes that not all points in an image have a high probability of being identified unambiguously during a comparison; rather, only certain points possess this property. These are notable, stable, and easily identifiable points.

In the last decade, as in nearly all fields of Computer Vision, significant advancements have been made in the development of local invariant features, characteristic points that enable applications to define a local geometry of the image and encode it in such a way that it remains invariant to image transformations, such as translation, rotation, scaling, and affine deformations.

In this chapter, we will address topics that are closely related to keypoint extraction algorithms. The discussion regarding point description will be covered in the following chapter, as it is an orthogonal topic to both the description of points and the concept of classification.

A non-exhaustive list of algorithms for keypoint detection includes:

Harris Corner

Harris mathematically formalizes the concept of an edge and, through the study of the eigenvalues of the covariance matrix in the vicinity of a point, allows for the determination of the presence or absence of a corner. It is invariant to changes in brightness, geometric transformations such as translations and rotations, and minimally to scale variations (see section 5.2);

KLT

The Kanade-Lucas-Tomasi method utilizes a variant of Harris (Shi-Tomasi) as a corner detector and performs matching using pyramid representations of the scene (details in 7.2);

AST

The class of Advance Segment Test (see section 5.5) identifies a characteristic point by observing the difference in brightness of points on a circumference;

SIFT

Analyzes the image in a multi-resolution manner and is invariant to similarity transformations (see section 5.3);

SURF

A more efficient variant of SIFT based on the integral image (see section 5.4).