首页 >
> 详细

COMP 558 Assignment 1

Posted: Thursday, Sept 23, 2021

Due: Thursday, Oct. 7, 2021 (by midnight, 11:59pm)

Introduction

This assignment covers material up to Lecture 6. The questions concern basic image processing using

convolution, and some of core ideas of edge detection, and RANSAC.

Please use the mycourses discussion board forum for assignment 1, for any assignment related ques-

tions. This will help us reach the entire class when we respond. However, follow appropriate protocol,

e.g., do not reveal the answer to a particular question or ask if your proposed solution is correct. Save

these kinds of questions for one-on-one interactions during office hours. Do not post pictures of your

solutions because these could give ideas to the rest of the class. You are free to discuss the general as-

pects of questions with each other, but not to the point where you are revealing answers or sharing solu-

tions. This is an individual assignment and the solutions that you submit must reflect your own work and

must be written by you. You are not permitted to submit code or text that you have copied from one

another or from any third party sources, including internet sites. By submitting this assignment you

agree to abide by these rules.

Instructions

Submit a single zipped file A1.zip to your mycourses Assignment 1 folder. The zip file must contain:

A single PDF file with figures and text for each question. Please do not spend time with fancy

typesetting. Just make sure that your explanations are clear and that the figures and their captions

are easy to interpret and that you answer each part of each question.

The Matlab code that you wrote so that we can run it if necessary. Make sure that you name the

files as specified, in case the TAs wish to run your code using their own script.

Any images that you used to test your code, in case we need to run your code to reproduce your

results.

In order to receive full points, your solution code must be properly commented, and you must provide a

clear and concise description of your computed results in the PDF. Note that the TAs have limited time

and will spend at most 20-30 minutes grading each submission.

Late assignment policy: Late assignments will be accepted up to only 3 days late and will be penalized

by 10% of the total marks, per day late. For example, an assignment that is submitted 2 days late and

that receives a grade of 90/100 would then receive an 18 point penalty. Submitting one minute after

midnight on the due date means it is 1 day late (not a fraction of a day), and same for the second day.

Tips

To examine the RGB values in an image, use the figure tab ToolsàDatatips. This will allow you to

click on pixels and see their values.

To speed up computation when testing code, resize the image (Matlab imresize).

Pay attention to the types of your data. You might need to use logical type (boolean), uint8, and

double. Matlab functions can be finicky and you may find you are spending much of your

debugging time on type errors. Another common error is to swap the x and y indices. If your

image is a 2D matrix, then you will index it with image(y, x) since matrices index row first and

column second. Also, since sometimes we will want to treat the centre of the matrix as the location

(0l,0), you might need to apply a shift in the (, ) coordinates for some of the steps below.

Make sure that you have the Computer Vision Toolbox and Image Processing Toolbox installed.

When you are working with “intensity” it is fine to convert an RGB image to a grayscale one, e.g.,

using rgb2gray

In Matlab think about how to vectorize computations and avoid loops if possible, e.g. = .?

creates a matrix which is the element-wise multiplication of two matrices , that have

compatible sizes.

Question 1 Marr-Hildreth [20 points]

a) Consider a scalar function (, ) of two variables and , where and are variables that are

associated with two orthogonal directions. The Laplacian of is given by ? = !!"!#! + !!"!$!, where we

assume that is at least twice differentiable. Show that the Laplacian is rotation invariant, i.e., the

value of the expression does not depend on the actual choice of the orthogonal directions.

[Hint: Consider an arbitrary orientation , with a variable representing position on a line along that

orientation. Then compute ? using calculus and the chain rule, but using this line as one of the

chosen coordinate directions. To do this properly you need to introduce a change of coordinates.]

b) Edge detection in 2D is in general computationally expensive because a priori we do not know the

local orientation of an edge. Hence, one has to look for zero-crossings of a second directional

derivative, taken in multiple local directions. This is in fact how the mammalian visual system

implements a form of edge detection in particular layers of the visual cortex. In their paper on edge

detection Marr and Hildreth get around this complexity by associating edge locations with zero-

crossings of the non-directional (as in rotation invariant) operator, the Laplacian. They claim that

this operation picks out the direction which has highest slope of the zero-crossing (if one were to

take directional derivatives that is). However, they make an important assumption about the local

intensity profile. What is that assumption? Explain briefly how, if that assumption holds, their

argument works.

c) Given your answer in part b) when would you expect the Marr-Hildreth edge detection strategy to

give good results and when you would expect it to fail? You can draw on your intuition from a

natural image of a complex scene, which you could present and use to defend your argument. You

don’t have to implement the strategy, you just have to provide your reasoning.

Question 2 Edge detection Using Oriented Operators [40 points]

Write a script Q2_Edgedetection.m, using Matlab, to answer the following questions.

a) In class we’ve seen 1D Gaussian’s and their derivatives. Let be an integer variable that is

associated with the standard deviation of a 1D Gaussian. Create a 2D array with dimensions

(2 + 1) x (2 + 1). The centre of the array represents the (, ) position (0,0). Now populate the

entries at each row in the array with the value of the second derivatives of a 1D Gaussian in the direction, with standard deviation . The 1D Gaussian should have its peak value in the middle of

the row, and the values in each row should be identical. The original Gaussian should also be

normalized to have area equal to 1 underneath it, prior to the computation of the second derivative.

Show an image of this filter and show the colormap using the colorbar command. Also visualize

the filter values as a 2D height function, using Matlab’s surface plotting functions (e.g., surf,

surfc, mesh, etc.). As an example, if you had a 2D Gaussian and you computed its partial

derivatives in the and directions, a visualization of the results would look something like the

following. (Ofcourse, the filter you are designing in this question is not the same thing.)

b) Now modify the filter above, by multiplying its entries by a normalized 1D Gaussian in the direction, with standard deviation $, with $ > . This should have the effect of tapering the filter

values in the direction. Show an image of this modified filter and show the colormap using the

colorbar command. If you like you can also visualize it using Matlab’s surface plotting functions

(surf, surfc, mesh, etc.).

c) You will now write a routine to create a rotated version of the 2D filter in part b). The goal is to

rotate it by a small amount . An effective way of doing this is just to rotate the coordinates. In

other words, for the value at a coordinate (&, ′), in the destination 2D filter, where these

coordinates represent a rotation of (, ) about the origin, we simply copy the value from the

source 2D filter, at location (, ). Show images of this modified filter for different =0, 45, 90 135 , including the colormap using the colorbar command. Also visualize

the rotated filters as 2D height functions, as you did for part b).

d) We have provided an image of skyscrapers in Manhattan, viewed under perspective projection.

Filter this image with your filter from c), but using the 4 rotated copies of it (0,45,90 and 135

degrees). For each copy, find the pixels at which there is a zero crossing in either the or or

diagonal directions. Adjust your choice of so that you get acceptable results. Create a binary

image that indicates where these zero crossing points are. Note: you will actually have 4 sets of

zero-crossings, one for each orientation of the filter so you can think of a clever way of presenting

these results together. One idea could be to show each set of zero-crossings in a different colour

and then use the Matlab functions such as imfuse to create a composite image. Alternatively, you

can use different colours for each filter orientation. You just have to explain your visualization.

For this part you have some flexibility in how you choose to display your results. You might need to

tune your parameters , $ . Do not use for loops to find the zero crossings. Instead, use

vectorization. Hint: You want to compare neighboring pixels of the filtered image and see where

the values have different sign, namely if their product is negative. To avoid the annoyances of

image boundaries, use Matlab’s circshift operation to compare an image to a shifted version of

itself. Note that logical operations can also be vectorized.

e) Repeat (d) but this time use a Laplacian of a Gaussian filter with = 6 and width 40. To construct

this filter you can directly use the Matlab function fspecial. Briefly discuss how your results in

e) compare with the results in d), using also your insights from question 1.

Question 3 Line finding using RANSAC [40 points]

For this question, you should use the Manhattan skyscraper image. You can also use an additional image

for illustrative purposes if you like, but make sure that it contains a certain distribution of long linear

structures. Don’t forget to submit this image as well.

a) Your task is to implement the RANSAC method to find prominent lines in your input image. The

pseudo-code for the RANSAC algorithm is in the slides and lecture notes provided in lecture 6, and it

is presented in slightly simplified form here. In this simplified form, initially we are not using least

squares to refine a fit, and we are not finding the `best’ model.

Repeat M times{

Randomly sample an edge element a list of N edge elements;

Examine all remaining (N-1) edge elements and count

the number C that are within some threshold distance

from the line associated with the edge element, and

whose orientation is within ? degrees of the orienta-

tion of the edge element.(This is the ‘consensus set’

for that line model.);

联系我们

- QQ：99515681
- 邮箱：99515681@qq.com
- 工作时间：8:00-21:00
- 微信：codinghelp

- 代写 Lab 2: Threads 2022-05-10
- 辅导assessment 1. Present Your Client ... 2022-05-10
- 5Cce2sas辅导、Python，Java程序辅导 2022-05-10
- 代写brae Webb编程 2022-05-09
- 辅导csci 3110 Assignment 1 2022-05-09
- Mth2222 Assignment 2代写 2022-05-09
- Cse3bdc Assignment 2022辅导 2022-05-08
- 辅导cis 468、辅导java，Python编程 2022-05-08
- Comp Sci 4094/4194/7094 Assignment 3 D... 2022-05-07
- Cs 178: Machine Learning & Data Mining... 2022-05-07
- Data7703 Assignment 4 2022-05-07
- 讲解assignment 2: Databases 2022-04-25
- 辅导ait681 Static Analysis 2022-04-25
- Cse121 & Cse121l 编程辅导、辅导c++程序语言 2022-04-25
- 辅导iti1120 Bject-Oriented Programming 2022-04-25
- Cmt304语言辅导、辅导c++，Python编程 2022-04-25
- 辅导comp/Engn4528 Computer Vision 2022-04-24
- 辅导fin 2200 Bloomberg Investment Proj... 2022-04-24
- 辅导bism 7255 Uml Assignment 2022-04-23
- 讲解comp202 Programming Assignment 2022-04-23