DD2423Note

DD2423Note

Posted by PY on September 10, 2024

DD2423

Lecture 1 Introduction to computer vision

What does it mean that vision is an active process?

image-20241210073011736

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

image-20241210073103179

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.

image-20241210073122801

Summary of good questions

• What is a pinhole camera model?

image-20241210154250723

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

image-20241210154259224

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

image-20241210154309193

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

image-20241210154317287

• How is a perspective projection expressed in homogeneous coordinates?

image-20241210154323533

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

image-20241210154330789

• What is an affine camera model?

image-20241210154337585

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

image-20241210154345849

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.

image-20241210073812334

image-20241210074256872

image-20241210074436889

Why are there so few real pinhole cameras?

image-20241210074503182

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?

image-20241210080445866

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

image-20241210080458816

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

image-20241210080514001

image-20241210080550922

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

image-20241210080604953

Lecture 3 Image filtering

What properties does a linear filter have? Multiple answers are possible.

image-20241210080641518

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

image-20241210082938171

image-20241210084143286

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

image-20241210084207674

img

Which Fourier transform corresponds to which image below?

image-20241210084253140

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

image-20241210091142580

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?

image-20241210091210758

image-20241210091348792

image-20241210091408105

image-20241210093002391

image-20241210091416794

image-20241210091427314

image-20241210091432404

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

image-20241210093046987

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

image-20241210093124278

image-20241210094455430

image-20241210094510790

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

image-20241210094539999

image-20241210094634038

Summary of good questions

• What properties does a linear shift-invariant filter have?

image-20241210154501136

• How do you define a convolution?

image-20241210154509994

• Why are convolutions important in linear filtering?

image-20241210154519455

• How do you define a 2D Fourier transform?

image-20241210154537425

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

image-20241210154550706

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

image-20241210154559163

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

image-20241210154606628

• What does separability of filters mean?

image-20241210154612885

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

image-20241210154620842

• How do you apply a discrete Fourier transform?

image-20241210154630794

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

image-20241210154637239

• In what sense is the Fourier transform symmetric?

image-20241210154643111

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.

image-20241210095027980

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)?

image-20241210095434010

image-20241210095703466

image-20241210095717729

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.

image-20241210095734675

Summary of good questions

• Mention a typical grey-level transformation. When to use it?

image-20241210154813885

• What do histogram stretching and compression mean?

image-20241210154829159

• What are the principles of histogram equalization?

image-20241210154841741

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

image-20241210154849815

• Why are ideal lowpass filter rarely used in practice?

image-20241210154902224

• What characteristics does a Gaussian filter have?

image-20241210154908500

• What is the difference between mean and median filters?

image-20241210154914422

• How can you do sharpening?

image-20241210154921244

• How can you approximate a first order derivative?

image-20241210154931088

• What is a Laplacian?

image-20241210154937622

Lecture 5 Image features I

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

image-20241210095858411

image-20241210093002391

image-20241210100321968

image-20241210100601481

image-20241210100632128

image-20241210100644402

image-20241210100651212

Summary of good questions

• Why do we get image aliasing when subsampling and what to do about it?

image-20241210155018278

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

image-20241210155026757

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

image-20241210155041706

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

image-20241210155051925

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

image-20241210155116885

• Why is edge detection important for image understanding?

image-20241210155126642

• What families of methods exist for edge detection?

image-20241210155136496

• What information do image gradients provide?

image-20241210155144847

• How does the Canny edge detector work?

image-20241210155151275

• What is differential edge detection?

image-20241210155157001

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

image-20241210155201778

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?

image-20241210100717522

image-20241210100800318

image-20241210100809912

image-20241210100822441

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:

  1. As you gradually increase the scale and blur the image, no new edges will appear,
  2. 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,
  3. The result of a Gaussian filter does not depend on the position in the image.

image-20241210100906389

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

image-20241210100927117

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

image-20241210100939675

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?

image-20241210100955162

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

image-20241210101031738

Summary of good questions

• What are the motivations for computing interest points? What are they typically used for?

image-20241210155333108

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

image-20241210155340134

• Why is scale selection an important operation?

image-20241210155346551

• Describe how scale selection can be performed in practice.

image-20241210155353257

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

image-20241210155358679

• How are image pyramids computed from image data?

image-20241210155403745

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

image-20241210155411103

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

image-20241210155416545

• How is the SIFT descriptor defined from image data?

image-20241210155421432

• How is the SURF descriptor defined from image data?

image-20241210155427376

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

image-20241210155447299

Lecture 7 Model fitting and representation

Which statements below regarding scale-space representations are true? Multiple correct answers might exist.

image-20241210102015088

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?

image-20241210140713504

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?

image-20241210140748616

image-20241210140937216

image-20241210140956364

image-20241210141014916

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

image-20241210141030628

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.

image-20241210141056767

Which statement is untrue regarding SIFT descriptors?

image-20241210141112618

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?

image-20241210141143207

image-20241210141504940

image-20241210141554857

image-20241210141533004

image-20241210141606838

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

image-20241210141623360

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.

image-20241210141658756

image-20241210141828790

image-20241210141841668

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

image-20241210141855229

image-20241210142010600

image-20241210142020265

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?

image-20241210141941351

Summary of good questions

• How does a Hough transform work for lines?

image-20241210155648407

• How many accumulators should you use?

image-20241210155653479

• How does RANSAC work?

image-20241210155657877

• What is a homography?

image-20241210155703513

• How do you create a connected component?

image-20241210155709430

• In what ways can two shape descriptors be different?

image-20241210155716049

• Can you give two examples of two shape descriptors?

image-20241210155722545

• How do you compute moment descriptors?

image-20241210155726560

• What do you like to preserve with dimensionality reduction?

image-20241210155732205

• How does a PCA work?

image-20241210155737245

Lecture 8 Deep networks for computer vision

What statements regarding deep learning for computer vision are true? Multiple correct statements are possible.

image-20241210142049015

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

image-20241210142115412

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

image-20241210142142353

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.

image-20241210142200822

Summary of good questions

• How does a perceptron work?

image-20241210155828261

• Why do you need an activation function?

image-20241210155838813

• What is a loss function?

image-20241210155849092

• How does backpropagation work in principle?

image-20241210155900818

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

image-20241210155907705

• What are the typical layers of a CNN?

image-20241210155914337

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

image-20241210155927439

• What is the structure of an autoencoder?

image-20241210155933071

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

image-20241210155938687

• What can GANs be used for?

image-20241210155942870

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

image-20241210155947183

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

image-20241210142314191

image-20241210142410595

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

image-20241210142442204

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?

image-20241210142528653

image-20241210142653767

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?

image-20241210142921405

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

image-20241210143048229

image-20241210143213549

image-20241210143222120

Summary of good questions

• What is the purpose of image segmentation?

image-20241210160018480

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

image-20241210160025178

• How does K-means work?

image-20241210160031731

• Why is spatial coherence important for segmentation?

image-20241210160037033

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

image-20241210160043257

• What is an affinity matrix?

image-20241210160050525

• What does Normalized Cuts try to optimize?

image-20241210160058666

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

image-20241210160105341

• What is the purpose of graph cuts for segmentation?

image-20241210160115896

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

image-20241210160120933

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

image-20241210160125195

• What does a morphological opening and closing operation do?

image-20241210160128819

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.

image-20241210143324063

image-20241210143543844

image-20241210143556159

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?

image-20241210143623256

image-20241210143644341

image-20241210143653348

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

image-20241210143708192

image-20241210143737696

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

image-20241210144221837

image-20241210144215118

image-20241210144237723

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

image-20241210144248091

Summary of good questions

• How does stereo work in general?

image-20241210160158420

• Why can you get double vision?

image-20241210160204399

• What is triangulation?

image-20241210160215551

• What is the relationship between disparities and depths?

image-20241210160222781

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

image-20241210160229274

• What are the key concepts of epipolar geometry?

image-20241210160239228

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

image-20241210160244744

• What might complicate stereo matching?

image-20241210160253808

• How does a simple stereo matcher work?

image-20241210160259004

• What constraints are often used to improve stereo matchers?

image-20241210160307749

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

image-20241210160311876

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?

image-20241210144313787

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?

image-20241210144338505

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.

image-20241210144357466

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

image-20241210144418562

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?

image-20241210144708546

bf83c910cd3d135e0bd1abaeaeb464b

Summary of good questions

• What is the difference between recognition and classification?

image-20241210160358994

• What makes a good feature space for recognition?

image-20241210160403072

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

image-20241210160408604

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

image-20241210160413478

• What does a typical feature based method consist of?

image-20241210160418501

• What steps does a Bag of Words approach include?

image-20241210160423413

• What are the dominating methods based on deep learning?

image-20241210160428287

• Why is object detection harder than object classification?

image-20241210160432134

• What characteristics does a nearest neighbour classifier have?

image-20241210160436478

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

image-20241210160442789

Lecture 12 Motion and optical flow

Why is motion estimation often considered harder than stereo? Multiple answers may be correct.

image-20241210144727940

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?

image-20241210144740532

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

image-20241210144751504

image-20241210144831627

image-20241210144844092

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?

image-20241210144905647

image-20241210144959216

image-20241210145012386

image-20241210145021247

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

image-20241210145031672

image-20241210145318812

Summary of good questions

• Why is motion more complex than stereo?

image-20241210160519774

• What is a Focus of Expansion?

image-20241210160526547

• What is Motion field and what is Optical flow?

image-20241210160532344

• How do you derive the optical flow constraint?

image-20241210160539416

• What is a Second moment matrix?

image-20241210160547070

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

image-20241210160552726

• What are the characteristics of Horn & Schunck?

image-20241210160558845

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

image-20241210160603952