Sift

Question 1

Scale normalization

Answer 1

Characteristic scale of blob by convincing it with Laplacian a at several scales to find max-response

Response decays as scale increases
Must be multiplied by theta^2 to keep in check

Question 2

Detection of scale summary

Answer 2

Scale-space kernel is Gaussian

Detect peaks with Gaussian pyramid

Difference of Gaussian with constant ration of scales is close in approximation to scale-normalized Laplacian

Question 3

SIFT

Answer 3

find local max/min of diff of gaussians over space and scale

Compute invariant feature descriptors

Question 4

SIFT algorithm

Answer 4

Construct sample space

LOG approx by DOG

locate DOG extrema

Key point localization

Assign orientation to key points

Generate key point descriptors (SIFT features)

Question 5

How to get rid of key points

Answer 5

Remove low contrast points

Remove after edge responses

Question 6

Orientation assignment

Answer 6

Create weighted histogram of local gradient directions in neighborhood of key point

Weights are
Gradient magnitude
Distance to key point
Theta = 1.5 * scale of key point

Question 7

Properties of sift

Answer 7

Can handle changes in viewpoint
Cane handle Illumination
Fast and efficient- can run in real time