UNIVERSITY COLLEGE LONDON
EXAMINATION FOR INTERNAL STUDENTS
MODULE CODE : COMP0137
ASSESSMENT : COMP0137A7UC / COMP0137A7PC
PATTERN
MODULE NAME : COMP0137 - Machine Vision
LEVEL: : Undergraduate (Masters Level) / Postgraduate
DATE : 11/05/2021
TIME : 10:00
This paper is suitable for candidates who attended classes for this
module in the following academic year(s):
Year
2020/21
Additional material
N/A
Special instructions
N/A
Exam paper word
count
N/A
TURN OVER
Machine Vision 1 TURN OVER
UCL Computer Science
Examination paper
Paper details
Academic year: 2020/21
Module title: Machine Vision
Module code: COMP0137
Exam period: Main summer examination period
Duration: 24 hours
Deliveries for
which intended:
A7P (taught postgraduate, level 7)
A7U (undergraduate, level 7)
Cohorts for
which intended:
2020/21; 2019/20
Instructions
You must answer ALL questions on ALL pages of this assignment.
A maximum of 100 marks is available. The marks available for each part of each question
are indicated in square brackets [n].
Submit your answers online through Moodle. Unless specified, you do not need to show
your work.
Machine Vision 2 CONTINUED
Section A: A Graphical Model
1. There is a Markov Model pictured below. Each discrete variable is illustrated to
show what values are possible. A "configuration" is a specific combination, where
each xi is in a specific state, for i=1..5. The unary costs, pairwise costs, and triplet
costs for each variable or combination of variables are specified in the tables
provided.
All variables are discreet, but they do not all have the same number of possible
states. For example, variable X1 can be “A” or “B”, while X2 can be “C”, “D”, or “E”.
The unary costs for all five variables are:
The pairwise and triplet costs are here:
Use the above information to answer the following questions.

x
1
x
2
x
3
x
4
x
5
Machine Vision 3 TURN OVER
(a) Configuration A is ACGJN. Compute the total cost of this configuration: _____
[7 marks]
(b) Configuration B is AEFKM. Compute the total cost of this configuration: _____
[7 marks]
(c) What is the Maximum Likelihood configuration? _ _ _ _ _
[7 marks]
(d) What is the Maximum a Posteriori configuration? _ _ _ _ _
[8 marks]
[Total for Question 1: 29 marks]
Section B: Solar Energy Task
2. Imagine your company develops solar-energy farms, for electricity generation.
Every year, you consider multiple empty fields throughout the UK, as possible
sites to install your simple solar panels, which are mounted flat, parallel to the
ground. Business, weather, and environmental factors drive you to build such
collection sites very selectively: you want to be fairly certain a field will have a
specific energy-yield before building an appropriately-sized set of solar panels
there.
Once panels are installed, it is easy to measure their minute-by-minute power
generation. But for practical reasons, assume it is very hard to install panels
prospectively. Instead, assume it is easy to install a wide-field-of-view dome
camera, to collect sky-images over long periods.
In your first year, you have three or four existing solar-energy farms, where you
can both measure the electricity yield, and snap corresponding dome-camera
photos. This is your initial source of training data. You will be training and
applying a compute vision model that infers energy yield from dome-camera
images.
[Question cont. over page]
Machine Vision 4 CONTINUED
[Question 2 cont.]
(a) Even without any energy-yield information, it's possible to train an auto-encoder
using the dome-camera images. Name one benefit of doing this.
Answer: _______
[6 marks]
(b) You can collect and analyze the paired data, so energy yield and corresponding
dome-camera images, every month. Why might it be unwise to make decisions just
from the data collected in January, the first month of operations?
Answer: _____________
[6 marks]
(c) You could train the model that predicts energy-yield using data in the dual
domain. Under what circumstances would this approach be preferable?
Answer: _____________
[6 marks]
(d) You could train a Maximum A Posteriori (MAP) or a Bayesian model. Describe one
example scenario where the difference between the two matters.
Answer: _____________
[6 marks]
(e) The potential sites, where initial filming using the dome-cameras has happened,
are laid out geographically on a map. The layout of these sites is distinctly not
grid-like. How would this affect the pairwise potentials, if you use a Markov
Random Field to regularize your energy-yield predictions?
Answer: _____________
[6 marks]
[Total for Question 2: 30 marks]
Machine Vision 5 TURN OVER
Section C: Miscellaneous Questions
3. There are pairs of probability distributions, where the second is conjugate to the
first. Now consider a situation where the second distribution IS NOT conjugate to
the first. Give one sentence for each, to describe a) one potential benefit of this
design choice to use non-conjugate-distributions, and b) one potential
disadvantage.
____________________________________________________________________________________________
____________________________________________________________________________________________
____________________________________________________________________________________________
____________________________________________________________________________________________
[6 marks]
4. We wish to determine the pose of a crash-test dummy, from a single input image.
There is exactly one dummy visible in the photo, and we'll assume it is not
occluded by anything else. We are given a simplified initialization algorithm,
where a limb-detector runs on the given input image. The detector detects a loose
bounding box around each separate limb of a crash-test dummy, e.g. a "limb" is
detected for the chest, left upper arm, the left lower arm, the left hand, etc.
However, the detector does not distinguish between these limbs, so each is just
one limb in an unordered list of detections.
(a) Using a maximum of two sentences, explain how would you practically obtain the
training data needed to determine the pairwise terms for a tree-model. For
context, you would use those terms to find the Maximum A Posteriori
configuration of limb-labels for all the visible parts of the crash test dummy,
replacing each generic "limb" label by its inferred body-part name.
____________________________________________________________________________________________
____________________________________________________________________________________________
[6 marks]
[Questions continued on next page]
Machine Vision 6 CONTINUED
(b) Now, instead of each label wi representing just the limb's body-part name, you
want to determine each limb's 3D position and 3D orientation. Using a maximum
of two sentences, describe why you might regret choosing dynamic programming
to find the MAP solution to this harder version of the problem? Use a 3rd sentence
to name an alternative and say what this alternative would sacrifice compared to
dynamic programming.
____________________________________________________________________________________________
____________________________________________________________________________________________
____________________________________________________________________________________________
[6 marks]
[Total for Question 4: 12 marks]
5. Find out what an "Event Camera" is. You may access Wikipedia, or any published
writing or manuals online. For each of the scenarios below, indicate whether it's
better to feed data to a computer vision algorithm from (choose one): a) A regular
CMOS camera, b) An Event camera, or c) Both cameras are equally effective in that
scenario.
5.a: Detect an animal sneaking around a garden at night. __________
5.b: Generate a deblurred colour wedding photo. __________
5.c: Feed a face recognition system for keyless entry to a secure building. __________
5.d: Inspect train tracks from a moving train. __________
5.e: Detect forgery in paper banknotes. __________
[12 marks total]
6. Give two distinct reasons why lines that are parallel to each other in the real world
can appear non-parallel in a photograph.
____________________________________________________________________________________________
[4 marks]
[Questions continued on next page]
Machine Vision 7 END OF PAPER
7. Your aim is to buy the right number of roof tiles, to replace an existing roof.
Imagine you record a colour image sequence while walking all the way around the
outside of a home. The house is standing in an open clearing, so no trees or other
objects block your view as you point the camera at it. You use the image sequence
to first compute a sparse 3D point cloud. Then you reconstruct a dense 3D point
cloud, and proceed to segment which 3D points belong to the roof. In one
sentence, describe the one critical step that is still needed to compute the surface
area of the existing roof.
____________________________________________________________________________________________
____________________________________________________________________________________________
[4 marks]
[Total for Question 3, 4, 5, 6, and 7: 41 marks]
[Total for all questions: 100 marks]