Feature Detection

cv

Feature is a kind of representation with some invariances between different images.

1 Point Feature Detection -- Harris

Plotting derivatives as 2D points

Fitting ellipse to each set of points

1.1 Harris Corner Detection

There are many different ways to describe corner reponse $R$, for example, $R=\lambda$, and here we use a more efficient way, which is the Harris Corner Detection.

How to distinguish corner and other features? We use the Corner Response Map belowed:

• R depends only on eigenvalues of H.
• R is large for a corner.
• R is negative with large magnitude of an edge.
• |R| is small for a flat region.

The Harris Corner Detection Algorithm Steps:

• step1: compute matrix $H$ for each pixel
• step2: compute $R=det(H)-\alpha*trace(H)^2$
• step3: select the points with $R$ is larger than the threshold and local maxima

Issues: the corners detected by Harris detector are no uniformly distributed in the image (more points locate in the area of high contrast)

1.2 Harris corner detection--improvement

Improvement: Adaptive non-maximal suppression (ANMS，自适应非极大值抑制), detect features that are both local maxima and whose response value is significantly (10%) greater than that of all of its neighbors within a radius r. (Brown, Szeliski and Winder, 2005).

1.3 Harris corner properties

• The ellipse shape keeps invariant after rotating an angle (eigenvalue keeps invariant).
  

• Illumination invariant
  Matrix H uses image gradient, linear illumination keeps unchange.

• Contrast change keeps partial invariant. But affect the number of detected features.
  

• Variant to the image scale
  

2 Point Feature Detection -- MOPS (Multiscale Oriented PatcheS descriptor)

Scale Invariant: To detect Harris corners in 5-level of image pyramid, the window represents the orientation and scale of feature point

Limitation: the images should have similar scale.

3 Point Feature Detection -- SIFT(Scale Invariant Feature Transform)

To detect and describe local features in an images.

• Invariant to scale change
• Invariant to rotation change
• Invariant to illumination change
• Robust to addition of noise
• Robust to substantial range of affine transformation
• Robust to 3D view point
• Highly distinctive for discrimination

We desire porperties for SIFT are:

• Position: Repeatable across different changes
• Scale: Automatic scale estimation

Intuition: find scale that gives local maxima of some function $f$ in both position and scale.

4 Coding Practice

Actually，Harris Corner is not always a good method for detecing corners, especially when processing complex images.

The other feature detector such as SURF, SIFT, ORB, is better than Harris Corner.