Page 1 of 5 [PTO]

COMP61342

Formative Assessment

UNIVERSITY OF MANCHESTER

SCHOOL OF COMPUTER SCIENCE

M.Sc. in Advanced Computer Science

Computer Vision

18th May 2020

Please answer ALL questions provided

Please submit a SINGLE PDF document with your answers

COMP61342

Page 2 of 5 [PTO]

1.

A student has a set of shapes extracted from a database of images, and wishes to study the

variation of shape across the database. An expert has already annotated each shape example

with suitable landmarks.

a) Why is it necessary to align a set of shapes before building a model of shape variation?

Describe in detail how such alignment could be done.

[5 marks]

After successful alignment, the student now has a dataset of shapes thus:

She now decides to apply Principal Component Analysis (PCA) to this dataset.

b) Explain in detail how PCA could be applied to this dataset.

How are the properties and output of PCA of use in the resultant statistical model of

shape?

[5 marks]

c) Explain in detail how the mathematical methods and statistical modelling methods

described above can be used to build a computer vision system suitable for finding a new

example of an object in a new image.

What are the main disadvantages of such a model-based vision system?

[10 marks]

End of Question 1

COMP61342

Page 3 of 5 [PTO]

2.

a) Briefly describe the main steps of performing image segmentation using mean-

shift clustering algorithm. [6 marks]

b) What are the advantages and disadvantages of

i. EM clustering algorithm [4 marks]

ii. Normalised cuts segmentation algorithm [4 marks]

c) Consider the data in figure 1.

Figure 1

What do you expect to happen if we run the K-means algorithm with two

clusters on this data set? Explain why you expect this to happen. [6 marks]

End of Question 2

COMP61342

Page 4 of 5 [PTO]

3.

Figure 2 shows a pair of stereo images from the surface of mars that have been

captured using a pair of calibrated cameras.

a) Explain what “calibrated cameras” means. [2 marks]

b) Define disparity in stereo vision. [2 marks]

c) Describe a method for detecting interest points in an image. [6 marks]

d) Explain how you could use the pair of images in figure 2 to calculate the

distances from the camera of the surface features that appear in the scene.

Figure 2

In your answer you need to consider all steps in the process, from images to

depth values. You also need to give a diagram to illustrate your answer.

[10 marks]

End of Question 3

COMP61342

Page 5 of 5 [PTO]

a) Give a brief outline of the three main constituents of a non-rigid pairwise image-

registration algorithm. Compare and contrast an algorithm that uses a non-parametric

representation of image warps, with one that uses a parametric representation.

[6 marks]

b) Describe, using diagrams or otherwise, a simple linear warping algorithm for defining

a one-to-one mapping between two triangulated meshes. You need only consider one

triangle of the entire mesh, and how to map to the corresponding point(s) in a second

triangle of another mesh.

[4 marks]

c) Outline at least three distinct applications of non-rigid image registration to

biomedical imaging, making clear in each case why registration is required/useful.

[6 marks]

d) Why might we wish to perform registration across an entire population? In such a

case, would fully groupwise registration be a better choice than repeated pairwise

registration? Give the reasons for your answer.

[4 marks]

End of Question 4

END OF EXAMINATION