9. Corner / Interest Point Usage

Corner / Interest Point:

• can help recognize a kind of object
• can be used in robotics
• can build 2D maps
• can build a panorama

• image alignment
• etc.

How to build parorama:

Purpose: match images

• Detect interest points / corner points in both images
• Find same interest points in both images
  • But the detection procedure should be done independently per image

Extract corner descriptors at interest points:

• Take some neighbourhood (small window) around each interest point in both images
• Take pixels in neighbourhood and compute a high dimensional vector
• Find invariance to geometric and photometric differences between the 2 views

Corner Feature:

Corners are image locations that have large intensity changes in more than one direction.

• The intensity change at a given pixel in the direction (u,v) is measured by sum-of-squared-difference
• (SSD) of all pixels in a nbhd of that window, and the associated pixel shifted by (u,v).

  Harris Corner Detection:

  

Eigenvalue Analysis:

Alternatives to using eigenvalues / Harris Detector:

• use
  • 
  • 
  • M is the matrix C before
  • k is a threshold on the value of R
  • A pixel with an R > threshold (experimental) is a corner

**

Invariance:

To decide which 2 pixels from 2 different images are actually the same point

• Translation
• Rotation in image plane
• Scale change
• Rotate out of camera plane (no good solution)

Instead of Harris, there are also other methods like SIFT/SURF matchings to detect invariances.

Differences between 2 methods:

• Harris features work only for some motions (rotation in camera plane, translation)
• SIFT/SURF features work for larger motions, and for different types of motions
  • blur, lighting, compression, all motion in the camera plane, and some motions out of the camera plane, etc.
  • but SIFT/SURF is slower

Reference:

• wikipedia
• handout of COMP4102: Introduction to Computer Vision from Carleton University School of Computer Science, 2019