SIFT

Question 1

What is a video in visual computing?

Answer 1

A sequence of still images (frames) shown over time to create the illusion of motion.

Question 3

What challenges does object recognition in video face?

Answer 3

Motion, lighting changes, scale and rotation, clutter, and background changes.

Question 5

What is an interest point in an image?

Answer 5

A pixel-level structure that is repeatable and distinctive, such as a corner or blob.

Question 7

What are desirable properties of an interest point?

Answer 7

Repeatable, distinctive, stable under transformations, subpixel accurate, and well-represented by a descriptor.

Question 9

What is a blob in image analysis?

Answer 9

A region with distinct intensity or texture that can be localized and measured at a certain scale.

Question 11

What is a scale space in image processing?

Answer 11

A set of images smoothed at increasing levels of Gaussian blur, used to detect features at different scales.

Question 13

What is the formula for scale space?

Answer 13

S(x, y, σ) = G(x, y, σ) * I(x, y), where G is a Gaussian filter and I is the image.

Question 15

How are blobs detected in scale space?

Answer 15

By finding local maxima or minima across both spatial and scale dimensions.

Question 17

What is the Difference of Gaussians (DoG)?

Answer 17

An approximation to the Laplacian of Gaussian, computed as the difference between two Gaussian-blurred images at different scales.

Question 19

Why is DoG used in SIFT?

Answer 19

It is computationally efficient and effective at detecting blobs at multiple scales.

Question 21

What is SIFT?

Answer 21

Scale-Invariant Feature Transform — a method to detect and describe distinctive features in an image.

Question 23

What are the main steps of SIFT?

Answer 23

Blob detection, keypoint localization, orientation assignment, descriptor creation, and normalization.

Question 25

Why is orientation assigned to a SIFT keypoint?

Answer 25

To make the descriptor invariant to image rotation.

Question 27

How is the principal orientation determined in SIFT?

Answer 27

By computing a histogram of gradient directions around the keypoint and choosing the peak.

Question 29

What is the formula for gradient orientation in SIFT?

Answer 29

θ = arctan(∂I/∂y ÷ ∂I/∂x)

Question 31

What does the SIFT descriptor represent?

Answer 31

A histogram of gradient orientations in spatial subregions around the keypoint.

Question 33

How many dimensions does a typical SIFT descriptor have?

Answer 33

128 dimensions (from 4x4 grid of 8-bin histograms).

Question 35

Why is the SIFT descriptor normalized?

Answer 35

To make it invariant to lighting changes and contrast.

Question 37

What is the L2 distance between descriptors used for?

Answer 37

To measure similarity; a smaller L2 distance indicates a better match.

Question 39

What is the formula for L2 distance between two histograms?

Answer 39

d(H₁, H₂) = sqrt(Σ(H₁(k) - H₂(k))²)

Question 41

What is histogram intersection used for in feature matching?

Answer 41

To measure similarity by summing the minimum values of corresponding bins in two histograms.

Question 43

What is the formula for histogram intersection?

Answer 43

Σ min(H₁(k), H₂(k))

Question 45

What is an application of SIFT in image alignment?

Answer 45

Matching keypoints between images to compute a homography for alignment or stitching.

Question 47

How is SIFT used in video tracking?

Answer 47

By matching features frame-to-frame to track the motion of objects.

Question 49

How does SIFT support object recognition in video?

Answer 49

By detecting and matching distinctive features invariant to scale, rotation, and lighting.

Question 51

What is the advantage of using descriptors like SIFT?

Answer 51

They are robust to transformations and help in reliable matching across images or video frames.