DD2423
Lecture 1 Introduction to computer vision
What does it mean that vision is an active process?

Given all other facts being equal, is the quality of a pixel in a high-resolution camera better than in a low-resolution camera?

Grey-level pixel values in digital cameras are usually stored in bytes (8 bits). Why can this be considered a serious limitation? There might be more than one correct answer among those below.

Summary of good questions
• What is a pinhole camera model?

• What is the difference between intrinsic and extrinsic camera parameters?

• How does a 3D point get projected to a pixel with a perspective projection?

• What are homogeneous coordinates and what are they good for?

• How is a perspective projection expressed in homogeneous coordinates?

• What is a vanishing point and how do you find it?

• What is an affine camera model?

• When to choose to a perspective model vs an affine one?

Lecture 2 Projections and transformations
Assume you have a point at (3,−2,8) with respect to the camera’s coordinate system. If the image has a size (w, h) = (640, 480), origin in the upper-left corner and the focal length is f = 480, what is the x-coordinate of the image projection? Give an answer without decimals.



Why are there so few real pinhole cameras?

If we assume we have a point in homogeneous coordinates (5, 3, 4), what is the x-coordinate of the same point in Euclidean coordinates?

A vanishing point is an image projection of some world point. Expressed in homogeneous coordinates, what is the last coordinate of this world point?

Which of the following alternatives includes a group of intrinsic camera parameters?


When can it be more convenient to use an affine camera model than a perspective one?

Lecture 3 Image filtering
What properties does a linear filter have? Multiple answers are possible.

What is the convolution of [1,2,3] and [1,2,3]?


What is the (magnitude of the) Fourier transform of a box function


Which Fourier transform corresponds to which image below?

What does the phase information of a Fourier transform contain? Multiple answers are possible.

Compute the discrete Fourier transform f (u) of the 4-pixel image f(u)=1, 3, 2, 1. Without normalisation, what is the magnitude of the Fourier transform?







Which ones of the following statements are true regarding Fourier transforms?

For a 1D signal, what does the conjugate symmetry and periodicity of the Fourier transform imply?



How do you compute the convolution h, = f* g of two discrete signals, f and g


Summary of good questions
• What properties does a linear shift-invariant filter have?

• How do you define a convolution?

• Why are convolutions important in linear filtering?

• How do you define a 2D Fourier transform?

• If you apply a Fourier transform to an image, what do you get?

• What information does the phase contain? And the magnitude?

• What is the Fourier transform of a convolution? Why important?

• What does separability of filters mean?

• How do you interpret a point in the Fourier domain in the spatial domain?

• How do you apply a discrete Fourier transform?

• What happens to the Fourier transform, if you translate or rotate an image?

• In what sense is the Fourier transform symmetric?

Lecture 4 Image enhancement
Assume that you apply the grey-level transformation below to an image. Which of the given statements are correct? There might be multiple correct answers.

Assume that you apply Histogram Equalization to an image with a grey-level distribution pz(z)= 2(1-z);z∈[0,1] using a transformation =T(z) € 0,1,where T(0)=0 and T(1)= 1.Which grey-level value z in the new image will the grey-level z = 0.5 be mapped to, that is what is T(0.5)?



Assume you do Histogram Equalization of a discrete grey-level image with 256 grey-levels to get a new image with an equal number of grey-levels on the output. Which of the following statements are true? There might be multiple correct answers.

Summary of good questions
• Mention a typical grey-level transformation. When to use it?

• What do histogram stretching and compression mean?

• What are the principles of histogram equalization?

• What are the differences between lowpass, bandpass and highpass filters?

• Why are ideal lowpass filter rarely used in practice?

• What characteristics does a Gaussian filter have?

• What is the difference between mean and median filters?

• How can you do sharpening?

• How can you approximate a first order derivative?

• What is a Laplacian?

Lecture 5 Image features I
Which statements below are true for the filter with kernel [1,-2, 1]? There might be multiple correct answers.







Summary of good questions
• Why do we get image aliasing when subsampling and what to do about it?

• Why is the notion of scale important in image analysis and computer vision?

• What is a scale-space representation? On what basis is it constructed?

• What structural requirements are natural to impose on early visual operations?

• What is meant by a Gaussian derivative? Why are these important?

• Why is edge detection important for image understanding?

• What families of methods exist for edge detection?

• What information do image gradients provide?

• How does the Canny edge detector work?

• What is differential edge detection?

• What should the image derivatives be equal to on edge points?

Lecture 6 Image features II
Which of the images below are the results of a Gaussian filter, a Median filter and an Ideal low-pass filter respectively, given the top-left input image?




air these scale-space properties that Gaussians satisfy
A. Semi-group property over scale,
B. Shift invariance over space,
C. non-enhancement of local extrema,
with what they mean in practise:
- As you gradually increase the scale and blur the image, no new edges will appear,
- If you apply two Gaussian filters with variances s1 and s2 in sequence, you get the same results as if you were applying one Gaussian filter with variance s1+s2,
- The result of a Gaussian filter does not depend on the position in the image.

Which ones of the following Gaussian operators are rotationally symmetric? Multiple answers might be correct.

What is true regarding Canny’s edge detector? Multiple answers might be correct.

For a differential edge detector, what are the conditions on the second order and third order derivatives in the gradient direction, i.e. Lvv and Lvvv respectively?

What happens during sampling if an image contains frequencies higher than half the sampling rate? Multiple answers are possible.

Summary of good questions
• What are the motivations for computing interest points? What are they typically used for?

• Describe three common interest point detectors including their mathematical definitions.

• Why is scale selection an important operation?

• Describe how scale selection can be performed in practice.

• What is the motivation for using image pyramids in computer vision?

• How are image pyramids computed from image data?

• Describe a basic trade-off issue that arises in hybrid pyramids.

• What is the purpose of computing image descriptors at interest points?

• How is the SIFT descriptor defined from image data?

• How is the SURF descriptor defined from image data?

• Outline the basic steps in an algorithm that matches interest points with associated image descriptors between two images of the same scene.

Lecture 7 Model fitting and representation
Which statements below regarding scale-space representations are true? Multiple correct answers might exist.

How can you tell based on the eigenvalues (λ1,λ2) of a second moment matrix computed in a patch around an interest point whether the interest point is good for matching?

When applying the Harris corner detector with constant k=0.05, a second moment matrix at a particular point is found to be
μ=(41)(12).
What is the Harris measure H at this point?




Which statements are true for the Determinant of Hessian, unlike the Laplacian, when applied for blob detection? Multiple answers might be correct.

What could you compute for blob detection in scale-space, if you want the scale t of where you find a local maximum in both scale and space to be indicative of the real size of the blob? Multiple answers might be correct.

Which statement is untrue regarding SIFT descriptors?

Assume that you have a shape given by the X-marks in the figure below, with an x-axis pointing to the right and an y-axis pointing downwards? What is the centered 2nd order moment x-wise of this shape?





Which statements mentioned below regarding Hough Transform are true? Multiple answers can be correct.

Assume that you use RANSAC to find straight lines. In the first round RANSAC randomly selects p1=(1,2) and p2=(3,5). Will the points p3=(2,2) and p4=(4,6) be regarded as inliers or outliers assuming the you measure errors in the y-direction and apply a threshold ϵ=1.



Which one of the following statements regarding homographies expressed as 3×3 matrices in homogeneous coordinates is false?



Assume that you do principal component analysis (PCA) on some data points in 2D. Given the points below, which directions are possible directions for the first principal component? Multiple answers might be correct?

Summary of good questions
• How does a Hough transform work for lines?

• How many accumulators should you use?

• How does RANSAC work?

• What is a homography?

• How do you create a connected component?

• In what ways can two shape descriptors be different?

• Can you give two examples of two shape descriptors?

• How do you compute moment descriptors?

• What do you like to preserve with dimensionality reduction?

• How does a PCA work?

Lecture 8 Deep networks for computer vision
What statements regarding deep learning for computer vision are true? Multiple correct statements are possible.

Which of the following statements is NOT true regarding activation functions?

Convolutional neural networks (CNNs) typically contain a number of layers of different types. Match the statements below to the correct layer type.
Makes CNNs gradually more translationally invariant.
Pooling layers
Uses weight sharing to reduce the number of parameters to train.
Convolutional layers
Can often be retrained for a quick transfer to new training data.
Fully-connected layers
Other Incorrect Match Options:
- Invariance layers
- Training layers

Over the years deep networks have gradually become better through a series of improvements. Which of the following statements are true regarding modern networks? Multiple statements may be correct.

Summary of good questions
• How does a perceptron work?

• Why do you need an activation function?

• What is a loss function?

• How does backpropagation work in principle?

• Why does a CNN have so many fewer parameters than a FCN has?

• What are the typical layers of a CNN?

• How do learned image descriptors perform compared to handcrafted ones?

• What is the structure of an autoencoder?

• What is the difference between a VAE and a GAN?

• What can GANs be used for?

• What are the basic components of a vision transformer network?

Lecture 9 Image segmentation
Assume that you want to apply thresholding to segment out foreground objects in three different images (A, B, C), where each image has it own distinct properties:
- A - a foreground with grey-level values clearly distinct from the background,
- B - a foreground for which you know the approximate size in the image, or
- C - the foreground and background gradually changes in grey-levels over the image.
Match the three cases to three possible thresholding methods:
- 1 - P-tile method
- 2 - Adaptive thresholding
- 3 - Mode method


Which statements regarding Mean-Shift segmentation are true? Multiple statements can be correct.

Let’s say we want to apply Normalized Cut to a group of 2x2 pixels [p1, p2 ; p3, p4] with links between neighbouring pixels using a 4-neighbourhood system. The links have weights w(p1,p2) = 3, w(p1,p3) = 2, w(p2,p4) = 1 and w(p3,p4) = 1. What is the minimal normalized cut Ncut in this case?


What is the most important reason why GrabCut normally works better than either Thresholding, Normalized Cuts and Mean-Shift for segmenting out a foreground object from its background?

Assume you apply a morphological opening operation on the image
| 0 | 1 | 1 | 1 | 1 |
|---|---|---|---|---|
| 0 | 0 | 1 | 1 | 1 |
| 1 | 1 | 1 | 1 | 0 |
| 1 | 1 | 1 | 0 | 0 |
| 1 | 1 | 1 | 1 | 1 |
with the structural element
| 0 | 1 | 0 |
|---|---|---|
| 1 | 1 | 1 |
| 0 | 1 | 0 |
What do you get?
After erosion you have
0 0 0 0 0 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 0 0 0 0 0
and dilation you have
0 0 0 1 0 0 0 1 1 1 0 1 1 1 0 1 1 1 0 0 0 1 0 0 0



Summary of good questions
• What is the purpose of image segmentation?

• How to select a threshold for histogram-based segmentation?

• How does K-means work?

• Why is spatial coherence important for segmentation?

• What does a mean-shift algorithm try to do and how?

• What is an affinity matrix?

• What does Normalized Cuts try to optimize?

• What cost functions does an energy formulation for segmentation often include?

• What is the purpose of graph cuts for segmentation?

• What kind of deep-learning based architectures exist for segmentations?

• How can you tell what object to segment with Segment Anything?

• What does a morphological opening and closing operation do?

Lecture 10 Stereo geometry
What is the depth (in meters) to a 3D point that has the projections (39, 54) and (24, 54) in two different parallel camera images with focal lengths f = 600 and baseline b = 0.10 m.



You are fixating at a point straight ahead with a vergence angle of 4 degrees. What is the depth of that point, assuming that the distance between your eyes is 6 cm?



For a stereo system with two cameras, what is an epipole? Multiple answers may be correct.


Assume you have a camera moving straight ahead along the optical axis. What the essential matrix between two consecutive images?



What is true regarding the monotonic ordering constraint used for stereo matching? Multiple answers may be correct.

Summary of good questions
• How does stereo work in general?

• Why can you get double vision?

• What is triangulation?

• What is the relationship between disparities and depths?

• Why does the error in depth increase for larger distances?

• What are the key concepts of epipolar geometry?

• What is an essential matrix and how is it used?

• What might complicate stereo matching?

• How does a simple stereo matcher work?

• What constraints are often used to improve stereo matchers?

• What parts do a deep network based stereo matcher often contain?

Lecture 11 Object recognition
Assume you have a model of a bicycle that contains a description of wheels, saddle, frame, chain, etc. Is model a discriminative model or a generative one?

Do the following questions represent a recognition, classification or a detection problem?
- A - What is shown in the center of this image?
- B - Is there a mobile phone somewhere in the image?
- C - Does this image show my mobile phone?

Let’s say we have an object recognition system based on matching of SIFT features. What kind of image changes is this system typically invariant to? Multiple answers may be correct.

Which statements are true regarding Bag of Words (BoW)? Multiple answers may be correct.

You have a basket of apples and bananas that are either green or yellow. Assume that 40% of them are bananas, 80% of bananas are yellow and 30% of apples are yellow. If you randomly pick a fruit that happens to be green, in percentage what is the probability that this is an apple?


Summary of good questions
• What is the difference between recognition and classification?

• What makes a good feature space for recognition?

• What kind of invariances do you often want in recognition?

• What classes of recognition methods exist and what are their differences?

• What does a typical feature based method consist of?

• What steps does a Bag of Words approach include?

• What are the dominating methods based on deep learning?

• Why is object detection harder than object classification?

• What characteristics does a nearest neighbour classifier have?

• How do you find a decision boundary for Bayesian classification?

Lecture 12 Motion and optical flow
Why is motion estimation often considered harder than stereo? Multiple answers may be correct.

By tracking a set of image points, you like to estimate the rotation (Rx,Ry,Rz) and translation (Tx,Ty,Tz) of a moving camera Assume that the x-axis is pointing to the left and the y-axis is pointing up. Which component is the easiest one to estimate?

What conditions are necessary for you to be able to compute the optical flow using the optical flow constraint?



Compared to Lucas & Kanade, does an affine motion model allow you to use larger windows around each point for which you like to determine the optical flow?




What is true regarding the monotonic ordering constraint used for stereo matching? Multiple answers may be correct.


Summary of good questions
• Why is motion more complex than stereo?

• What is a Focus of Expansion?

• What is Motion field and what is Optical flow?

• How do you derive the optical flow constraint?

• What is a Second moment matrix?

• How does Lukas & Kanade cope with large optical flow?

• What are the characteristics of Horn & Schunck?

• How do modern methods avoid problems around the borders of objects?
