Unit 4 - Practice Quiz

1 What is a local feature in the context of computer vision?

2 What type of image structure is the Harris detector primarily designed to find?

3 What does the acronym SIFT stand for in computer vision?

4 The SURF algorithm was primarily designed as a faster alternative to which popular feature detector?

5 What is the primary advantage of using binary feature descriptors over floating-point descriptors?

6 How does the FAST algorithm typically identify a corner?

7 BRIEF creates an image descriptor by performing what specific operation?

8 ORB was introduced as a free and efficient alternative to SIFT and SURF. What two algorithms does it combine?

9 What does HOG stand for in computer vision?

10 Which of the following is a common real-world application of feature detection and description?

11 What is the primary goal of feature matching?

12 Which similarity measure is computationally best suited for comparing binary feature descriptors like ORB and BRIEF?

13 How does a Brute Force matcher find the best match for a feature from a first image?

14 In the context of feature matching, what is a K-D tree primarily used for?

15 What is the core idea behind Locality-Sensitive Hashing (LSH)?

16 What does RANSAC stand for?

17 In the context of feature matching, what is the main purpose of the RANSAC algorithm?

18 Which of the following properties makes SIFT highly reliable for matching images taken from different distances and angles?

19 Which distance metric is most commonly used to match continuous floating-point descriptors like SIFT and SURF?

20 HOG descriptors were traditionally and most famously popularized for which specific computer vision task?

21 Which of the following characteristics is most crucial for a local feature detector to ensure that the same physical point in a scene is extracted across images taken from different viewpoints?

22 In the Harris corner detector, the structure tensor (second-moment matrix) is computed. Let and be its eigenvalues. What does it indicate if and ?

23 The Harris response function is given by . Why is this specific formula used instead of directly computing the eigenvalues?

24 How does the Scale Invariant Feature Transform (SIFT) algorithm initially identify potential keypoints across different scales?

25 To achieve rotation invariance, SIFT assigns a dominant orientation to each keypoint. How is this dominant orientation determined?

26 SURF accelerates the feature detection process compared to SIFT. What mathematical tool does SURF heavily rely on to speed up its filtering operations?

27 The FAST detector examines a circle of 16 pixels around a candidate keypoint. To quickly reject non-corners, which pixels are typically tested first?

28 How is a BRIEF descriptor formed for a given keypoint patch?

29 Standard FAST keypoints lack orientation, and standard BRIEF descriptors are not rotationally invariant. How does ORB modify FAST to compute a keypoint orientation?

30 In the HOG (Histogram of Oriented Gradients) descriptor, what is the primary purpose of applying block normalization across overlapping blocks?

31 When building an object recognition system for a heavily cluttered scene, why are local features (like SIFT) preferred over global features (like global color histograms)?

32 Which distance metric is computationally optimal for matching binary descriptors such as BRIEF, ORB, or BRISK?

33 What is the primary objective of Lowe's ratio test during the feature matching phase?

34 Given an image with descriptors and an image with descriptors, what is the time complexity of brute force matching if we want to find the best match in for every descriptor in ?

35 While K-D trees are often used to accelerate nearest neighbor searches in feature matching, under what condition does a K-D tree's performance typically degrade to that of a linear (brute force) search?

36 What is the core principle behind Locality-Sensitive Hashing (LSH) for approximate nearest neighbor search?

37 In RANSAC, let be the probability that a selected match is an inlier. If a model requires matches to be computed, what does the expression represent?

38 When using RANSAC to compute a homography matrix between two images, what is the minimum number of point correspondences () required to instantiate a model hypothesis in a single iteration?

39 What happens to the required number of RANSAC iterations if the proportion of outliers in your feature matches increases dramatically, assuming you want to maintain a 99% confidence of finding a good model?

40 Binary descriptors are primarily favored in real-time computer vision applications (like mobile AR) because they are memory efficient and fast to match. Which of the following is NOT a binary descriptor?

41 The Harris corner detector computes a corner response function . If the eigenvalues and of the second-moment matrix satisfy (where is very large), how does the Harris corner response behave, and what does it indicate structurally?

42 In the SIFT algorithm, extreme points are detected in the Difference of Gaussian (DoG) scale space. During sub-pixel localization, the Taylor expansion of the DoG function is used. If the calculated offset has a component larger than $0.5$ in any dimension, what action is taken?

43 SURF achieves computational efficiency by using Box Filters to approximate the Hessian matrix. How does SURF evaluate these Box Filters independent of the filter scale in time?

44 The FAST feature detector evaluates a circle of 16 pixels around a candidate pixel . To accelerate the rejection of non-corners, FAST initially examines a specific subset of pixels. Which subset is typically checked first to quickly reject a non-corner candidate?

45 BRIEF descriptors are highly sensitive to in-plane rotation. How does the ORB algorithm modify the BRIEF descriptor extraction to achieve rotation invariance (steerable BRIEF)?

46 In HOG descriptor computation, block normalization is a critical step for contrast invariance. If L2-Hys normalization is applied to a block vector , which of the following best describes the process?

47 When using RANSAC to estimate a model from a set of putative point matches containing an unknown outlier ratio , the number of iterations required to ensure a probability of picking at least one outlier-free sample of size is given by:

48 In the context of approximate nearest neighbor search using LSH for binary descriptors like ORB, which hash function family is commonly utilized to preserve the Hamming distance?

49 Why does the standard K-D tree approach degrade to linear search when matching high-dimensional descriptors like the 128-dimensional SIFT?

50 When matching two sets of feature descriptors using the Sum of Squared Differences (SSD) versus Normalized Cross-Correlation (NCC), which of the following conditions strictly favors NCC over SSD for robust matching?

51 A common strategy to eliminate ambiguous feature matches is Lowe's ratio test. If the nearest neighbor distance is and the second nearest neighbor distance is , the match is kept if . What underlying assumption justifies this test in complex scenes?

52 ORB uses rBRIEF (rotation-aware BRIEF) for its descriptor. To reduce the correlation among the binary tests and maximize variance, ORB applies a specific learning step during its design. What method is used to select the optimal pairs of pixels for the binary tests?

53 Which mathematical concept provides the theoretical foundation for treating feature detection as finding locations where the signal changes significantly in multiple directions?

54 In bag-of-visual-words (BoVW) image classification, visual vocabularies are created by clustering feature descriptors. If an image has multiple repetitive structures yielding many identical descriptors, how does the Term Frequency-Inverse Document Frequency (TF-IDF) weighting scheme handle these specific visual words?

55 SIFT handles rotation invariance by computing a dominant gradient orientation for each keypoint. If a local region has multiple peaks in the orientation histogram that are within 80% of the highest peak, how does SIFT process this?

56 In PROSAC (Progressive Sample Consensus), an extension of RANSAC, how is the sampling strategy modified to achieve faster convergence in feature matching?

57 FAST uses a machine learning approach to generalize and speed up corner detection. How is the decision tree formulated in the machine learning version of FAST?

58 In the formulation of the Difference of Gaussian (DoG) function , the DoG provides an approximation to the scale-normalized Laplacian of Gaussian (LoG). What is the theoretical relationship between the DoG and the scale-normalized LoG ?

59 Binary descriptors like BRISK rely on a specific sampling pattern around the keypoint. Unlike BRIEF, which uses random point pairs, what defines the spatial sampling pattern of BRISK?

60 When applying Brute Force matching between image A ( features) and image B ( features) using cross-check (mutual consistency) validation, a match between feature and feature is retained only if: