SIFT and Feature Matching

Question 1

What is scale invariance?

Answer 1

That laws dont change depending on scale. So a human in the distance is still a regular human, not a tiny human.

Question 2

What is the scale space?

Answer 2

A way to represent images with scale invariance. You can use sequential gaussian blurring to represent how objects are further away. Octave is the level of scale.

Question 3

What is keypoint localization?

Answer 3

A step in SIFT.
You need to detect local max/min in the DoG space. If a pixel is max or min, check in a 3x3x3 cube around the pixel. Only consider points that have neighbors above and below.
The you eliminate points with low contrast values, and eliminate edge responses.

Question 4

How can you compensate for orientation?

Answer 4

You calculate gradient magnitude and direction for each point. You then create a HOG for each point, which gets put into bins in 360 degrees. Choose only the responses above 80 percent.

Question 5

What is the sift descriptor then mate?

Answer 5

SIFT does several things:

• Detect scale space
• Localize Keypoints
• Detect orientation (36 bins for 360 degrees)
• Make keypoint descriptor (16x16 into 4x4)
• Match keypoints

Question 6

What is feature matching and how can it be done?

Answer 6

Given image 1, how to find best match in image 2?
You can define a distance function (normal one is Euicledian). Next you test against all found features in image 2, and get the ones with min distance/max similarity.

Question 7

What are the different feature matching-methods, and which is best?

Answer 7

• Sum of square differences: You use SSD to find distances

- Ratio of SSD is better: Look at ratio between different matches. Lower is better

Question 8

What is a ROC Curve?

Answer 8

Measures performance in matching, best results are as close to upper left corner as possible. Plots the TP vs FP rate.