COSC 445 – Computer Vision
Winter 2020 T1
Dr. M. S. Shehata

A7 (20 points) – Due Nov. 27th, 2020
Focus: Corner Detection, HOG, SVM
1. [10 points]: Implement corner detection algorithm using the following steps:
a) Filter the image f(x, y) with derivatives of a Gaussian G(x, y, σ1). You can use Sobel operator to
apply the derivative. For Guassian use mask size 6 σ1
(, ) = (, , 1) ∗
′(, ) = ∗
′(, ) = ∗
b) Form three maps:
(, ) = (′)
2
(, ) = (′)(′)
(, ) = (′)
2
c) Blur these maps with a Gaussian G(x, y, σ2), with mask size 6 σ2:
(, ) = (, , 2) ∗
(, ) = (, , 2) ∗
(, ) = (, , 2) ∗
d) Compute:
(, ) = +
(, ) = −
2
e) Compute:
(, ) = − 2
f) Compute local max of R(x, y) by applying non-maximum suppression over a 3 × 3 neighborhood
of each point.
g) Threshold R0 is used to prune points with R(x, y) < R0, where R0 = 0.01×max(R(x, y))
h) Locate all points in R(x,y) and draw small rectangle at each point on the input image

Write a Matlab function DetectCorners that detect the corners in a given image. Then display the
detected corners over the image
I = imread('4.png');
ImageWithCorners = DetectCorners(I,0.7,3,0.04);

The first parameter is the input image, second and third parameters are σ1, σ2 and the last
parameter is α
COSC 445 – Computer Vision
Winter 2020 T1
Dr. M. S. Shehata



[10 points]: Write a Matlab script that extract Histogram of Oriented Gradients (HOG) features from the
training images, and use the extracted features to train Support Vector Machine (SVM). Then use the
trained SVM to classify the testing images to either Human or Not Human. In you implementation, after
reading each image you need to resize it to a fixed size (128,64) to get a fixed size feature vector from HOG.
To extract HOG you can use the matlab built-in function “extractHOGFeatures”. Training of
the SVM using the function “fitcsvm” and to classify use the function “predict”.

Create a table in .docx file showing the classification accuracy versus
different cell size (2, 4, 8, 16, 32) for HOG

.





Submission Instructions
1- For question Q1 write a separate Matlab program or function. For question 2, upload the
Matlab code and .docx file with the table of accuracies. These two questions are to be
solved individually (no group work).
2- Submit everything as one zip file to Canvas. Note that you can resubmit an assignment, but
the new submission overwrites the old submission and receives a new timestamp.
Marking guide:
+8 for all steps (i.e., +1 for each step)
+1 displaying corners
+1 logical flow and correctness
Marking guide:
+1 resize images into fixed size
+2 extracting HOG features
+2 training SVM
+2 classifying using SVM and
calculating accuracy
+1 reaching accuracy >= 80%
+2 table of accuracies in .docx file

欢迎咨询51作业君