This my presentation about SIFT features at Sharif University of technology, Tehran, Iran. This presented in Machine Vision Course offered by Dr. M.Jamzad.
The presentation contains animations and it can not play properly! Please send e-mail to get the original one: sinaee@ce.sharif.ir
8. Page 8
Searching over all scales in order to identify the Location and Scales that can
be assigned under differing views of a same object.
To efficiently detect stable keypoint locations in scale space, Lowe(1999) use
DoG of two nearby scales,
11. Page 11
Finding the minimum or maximum sample point
among its 26 neighbors
The extrema may be close to each other and it cause
to be quite unstable to small perturbations of image
This problem arises from the frequency of samples
being used for detection of extrema.
Unfortunately, there is no minimum spacing of samples to detect all extrema
14. Page 14
Once keypoint candidates has been found, we want to reduce the response to the low
contrast points, or poorly localized along an edge
If the extremum is greater than 0.5 it means the extremum is closer to another
sample point.
15. Page 15
The value of the extremum is useful to reject the unstable extrema with low contrast.
Original Image Keypoints from extremas of DoG, 832Keypoints 729,
after threshold on the minimum contrast
19. Page 19
Peaks in histogram shows dominant directions in the spatial domain.
Highest peak and any one in the 80% of it are used to create a keypoint orientation.
For those who have the multiple peak of the same magnitude, there will be multiple
keypoint at a same point and location but different orientation.
20. Page 20
As it can be seen that SIFT is robust to
image noises
78% repeatability
10% of image pixel noise
27. SIFT keypoints are useful due to their
distinctiveness for object detection.
They are invariants to scale, orientation,
affine transformation.
They are robust to clutter backgrounds.
Page 27