Brainscape's Knowledge GenomeTM

Browse over 1 million classes created by top students, professors, publishers, and experts.


See full index

KCL computer vision > w6 gemini > Flashcards

w6 gemini Flashcards

1
Q

What is the goal of matching in computer vision?

A

To find corresponding points or regions between two or more images.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

Why is searching required in matching?

A

To find the ‘most similar’ points between images, whether using correlation-based or feature-based methods.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

What are ‘putative matches’?

A

The points initially identified as the most similar between images before outlier removal.
Could be inlier or outlier

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

What are ‘inliers’ in the context of matching?

A

Correct putative matches.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

What are ‘outliers’ in the context of matching?

A

Incorrect putative matches.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

What is the challenge in finding true correspondence between images?

A

The presence of matching errors (outliers).

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

What is the goal when dealing with outliers in matching?

A

To estimate the true transformation between images despite erroneous correspondences.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

What are the steps to find the most likely transformation despite outliers?

A
  1. Extract features (if using feature-based). 2. Compute putative matches. 3. Find most likely transformation (highest inliers, fewest outliers).
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

What algorithm is commonly used to find the most likely transformation in the presence of outliers?

A

RANSAC (RANdom SAmpling & Consensus).

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

What is the objective of the RANSAC algorithm?

A

To robustly fit a model to a dataset containing outliers.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q

What are the requirements for using RANSAC?

A
  1. Data consists of inliers and outliers. 2. A parameterized model explains the inliers.
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

What are the steps in the RANSAC procedure?

A
  1. Randomly choose a minimal subset of data points.
  2. Fit the model to this subset.
  3. Test all other data points for consistency with the model.
  4. Count the number of inliers (consensus set).
  5. Repeat for N trials. After N trials, select the model with the highest support. (Support is the cardinality of inliers set)
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

What is a ‘minimal subset’ (or sample) in RANSAC?

A

The smallest number of data points required to estimate the model parameters.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
14
Q

What does it mean for a data point to be ‘consistent’ with the fitted model in RANSAC?

A

The data point lies within a certain distance (threshold t) of the model’s prediction.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
15
Q

What is the ‘consensus set’ in RANSAC?

A

The set of data points that are consistent with the fitted model.
Aka the set of inliers

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
16
Q

What happens after N trials in RANSAC?

A

The model parameters with the highest support are selected, and the model can be re-estimated using all the points in this subset.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
17
Q

In the simple correspondence example, what is the model being fit?

A

A pure translation between the two images.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
18
Q

How many putative matches are needed to define a pure translation in 2D?

A

One.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
19
Q

In the simple correspondence example, what happens when a randomly chosen match is an outlier?

A

The fitted translation will likely be incorrect, leading to few or no other inliers.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
20
Q

In the simple correspondence example, what happens when a randomly chosen match is an inlier?

A

The fitted translation will likely be correct, leading to more other inliers.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
21
Q

What is the role of the ‘consensus set’ size in RANSAC?

A

It represents the ‘support’ for a particular model hypothesis. Larger consensus sets indicate a more likely correct model.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
22
Q

How does RANSAC handle more complex transformations?

A

By sampling more pairs of points (e.g., 4 pairs for a homography).

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
23
Q

What is a common method to extract interest points for matching?

A

Harris corner detector.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
24
Q

What is SSD used for in the context of the real correspondence example?

A

To find the best match for each interest point within a search window.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
25
Q

What does a line between interest points in the real correspondence example represent?

A

A putative match.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
26
Q

What is the outcome of applying RANSAC to the real correspondence example?

A

It identifies a model consistent with a large number of matches (inliers) and rejects inconsistent matches (outliers).

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
27
Q

What is a key advantage of RANSAC in real-world scenarios?

A

It can find correspondence even with a high number of outliers.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
28
Q

Besides finding correspondences, what else can RANSAC be used for?

A

Fitting algorithms, such as fitting a straight line to a set of points.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
29
Q

What is another algorithm for fitting a model to data besides RANSAC?

A

Hough Transform.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
30
Q

In the line fitting example, what is the model?

A

A straight line.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
31
Q

How many data points are needed to fit a straight line?

A

Two.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
32
Q

In the line fitting example, what happens when the initially chosen points are outliers?

A

The fitted line will not represent the majority of the data points.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
33
Q

In the line fitting example, what does testing other data points against the fitted line determine?

A

Whether those points are inliers or outliers for that specific line hypothesis.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
34
Q

What are the advantages of RANSAC?

A

Simple and effective, general method for various model fitting problems (segmentation, camera transformation, object trajectory).

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
35
Q

What are the disadvantages of RANSAC?

A

Requires many iterations if the percentage of outliers is high, lots of parameters to tune.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
36
Q

What is the correspondence problem in summary?

A

Finding matching image elements across images.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
37
Q

Where does the correspondence problem arise?

A

Stereo vision, video analysis, object recognition.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
38
Q

How is correspondence similar to grouping?

A

Grouping looks for similar elements in a single image, while correspondence looks for the same elements in multiple images.

39
Q

What are the two main approaches to solving the correspondence problem?

A

Correlation-based methods and feature-based methods.

40
Q

What is the core idea of correlation-based methods?

A

Matching image intensities, usually over a window of pixels.

41
Q

What is the core idea of feature-based methods?

A

Matching sparse sets of image features.

42
Q

What are the design decisions involved in solving the correspondence problem?

A

Which locations to match, what properties to match, where to look for matches, how to evaluate matches, how to find true correspondence.

43
Q

What are some examples of properties to match in correspondence?

A

Image intensities, descriptors of image properties (e.g., SIFT).

44
Q

What are the two approaches for selecting image locations to match?

A

All locations (correlation-based) or selected interest points (feature-based).

45
Q

What are the two approaches for where to look for matches?

A

Exhaustive search or restricted search.

46
Q

How can matches be evaluated?

A

Using similarity measures (correlation, normalized correlation) or difference measures (SSD, SAD).

47
Q

What is the core idea of feature-based methods for matching?

A

Matching based on a sparse set of features within each image.

48
Q

What are the steps in feature-based matching?

A

Detect interest points, find corresponding pairs of points by comparing features.

49
Q

What are the advantages of feature-based methods?

A

Relatively insensitive to illumination and size changes, less computationally expensive than correlation-based methods.

50
Q

What are the disadvantages of feature-based methods?

A

Provides a sparse correspondence map, only suitable when good interest points can be extracted.

51
Q

What are the requirements for good interest points?

A

Repeatable detection and distinctive descriptors.

52
Q

What does a repeatable detector ensure?

A

The same point is detected independently in both images, even with changes in scale, rotation, translation, and illumination.

53
Q

What does a distinctive descriptor ensure?

A

Corresponding points can be correctly matched with high probability.

54
Q

What are corners as interest points?

A

Points where two edges meet, characterized by high intensity gradients in two directions.

55
Q

What are eigenvalues of the Hessian matrix related to?

A

The maximum slope of intensity gradient at two orthogonal directions.

56
Q

What is the condition for a corner based on eigenvalues?

A

Both eigenvalues are large.

57
Q

How does the Harris corner detector avoid calculating eigenvalues?

A

By defining a measure R based on the determinant and trace of the Hessian matrix.
R = det(H) -k*(Trace(H))^2

58
Q

What are the characteristics of R for a corner, edge, and flat region in the Harris detector?

A

Corner: R is large and positive.
Edge: R is negative with large magnitude.
Flat: |R| is small.

Therefore corner is where R > threshold

59
Q

What is the final step in the Harris corner detector to identify interest points?

A

Taking the points of local maxima of R after thresholding.

60
Q

What is non-maximum suppression?

A

A process of setting the R value of a pixel to 0 if it has a neighbor with a larger R value.

61
Q

Why is non-maximum suppression used?

A

To ensure that only the strongest corner responses in a neighborhood are selected.

62
Q

What are the steps in the Harris corner detector algorithm?

A

Compute derivatives
compute products of derivatives
compute sums of products
define the Hessian matrix
compute the detector response R = Det(H) - k*(Trace(H))^2
threshold and apply non-maximum suppression.

63
Q

What is the impact of translation and rotation on Harris corner detector results?

A

It is invariant to translation and rotation. Eigenvalues remain the same.

64
Q

What is the limitation of the Harris corner detector regarding scale?

A

It is not scale invariant.

65
Q

How can scale invariance be addressed in corner detection?

A

By performing corner detection across a range of scales using an image pyramid (Harris-Laplacian).

66
Q

What is another scale-invariant interest point detector?

A

Scale Invariant Feature Transform (SIFT).

67
Q

What is the key difference in interest point detection between Harris-Laplacian and SIFT?

A

Harris-Laplacian uses the Harris corner detector in space and scale, while SIFT uses the Difference of Gaussians.

68
Q

What is the Difference of Gaussians (DoG) used for in SIFT?

A

To detect interest points in scale space.

69
Q

How are maxima and minima of DoG selected as interest points?

A

If a pixel’s value is larger or smaller than all its neighbors in a 3x3x3 neighborhood in scale space.

70
Q

What additional steps are involved in SIFT interest point detection after finding DoG extrema?

A

Keeping points with high contrast and sufficient structure using a threshold based on the ratio of trace and determinant of the Hessian matrix.

71
Q

What is required to find corresponding pairs of points after detecting interest points?

A

A measure of similarity between the points (a descriptor).

72
Q

What is the descriptor used in the basic Harris method?

A

A small window around the interest point (pixel intensity values).

73
Q

What similarity measures can be used with the basic Harris descriptor?

A

Euclidean distance, SSD, SAD.

74
Q

What are the limitations of the basic Harris descriptor?

A

Not robust to rotation, scale, changes in viewpoint or illumination.

75
Q

What is the descriptor used in SIFT?

A

A 128-element vector of intensity gradient orientations around the interest point.

76
Q

How is the SIFT descriptor created?

A
  1. Calculate orientation and magnitude of intensity gradient. 2. Create a histogram of orientations. 3. Create separate histograms for sub-windows.
77
Q

How is robustness to rotation achieved in the SIFT descriptor?

A

By rotating all orientations based on the dominant orientation.

78
Q

How is the final SIFT descriptor represented?

A

A 128-element vector, normalized to unit length.

79
Q

What similarity measure is commonly used with the SIFT descriptor?

A

Euclidean distance between vectors.

80
Q

What are the robustness properties of the SIFT descriptor?

A

Robust to translation, rotation, scale, changes in viewpoint and illumination.

81
Q

What is the basic idea behind correlation-based methods for matching?

A

Matching pixel values within image regions.

82
Q

What is the search region in correlation-based matching?

A

The area in the second image where the corresponding region is searched for.

83
Q

How is similarity computed between regions in correlation-based matching?

A

Using measures like cross-correlation, normalized cross-correlation, or correlation coefficient.

84
Q

What are the decisions to make when using correlation-based methods?

A

Size of correlation window and search area, and the method to measure similarity.

85
Q

What are the trade-offs for a small correlation window?

A

May not capture enough structure, may be noise sensitive.

86
Q

What are the trade-offs for a large correlation window?

A

Decreases precision, decreases tolerance to viewpoint.

87
Q

Why is the size of the search area important?

A

Full correlation is computationally expensive.

88
Q

How is the search area typically constrained?

A

Arbitrarily around the original pixel location or using task-specific knowledge (e.g., epipolar geometry).

89
Q

What are some common similarity measures in correlation-based methods?

A

Cross-correlation, normalized cross-correlation, correlation coefficient, SSD, SAD, Euclidean distance.

90
Q

Which similarity measures should be maximized?

A

Cross-correlation, normalized cross-correlation, correlation coefficient.

91
Q

Which similarity measures should be minimized?

A

SSD, Euclidean distance, SAD.

92
Q

Why is SAD often used in practice despite other measures?

A

It is simple and computationally efficient, and the performance difference is often negligible.

93
Q

What are the advantages of correlation-based methods?

A

Easy to implement, provides a dense correspondence map.

94
Q

What are the disadvantages of correlation-based methods?

A

Computationally expensive, needs images with distinct patterns, doesn’t work well with viewpoint changes or illumination changes.

KCL computer vision (12 decks)

Brainscape helps you reach your goals faster, through stronger study habits.
© 2025 Bold Learning Solutions. Terms and Conditions