Roll number:
Name:
MIDTERM EXAM
CSCS 457: COMPUTER VISION
Total Exam Time: 50 minutes
Instructions: Switch off your cell phone.
Problem 1 [8 pts]
Given the following 5×5 image patch and a 3×3 Laplacian mask (edge
detector):
Patch (intensity values):
[ ]
10 10 20 10 10
10 15 25 15 10
20 25 40 25 20
10 15 25 15 10
10 10 20 10 10
Laplacian mask (often used for edge detection, responds to intensity
changes in all directions):
[ ]
0 −1 0
−1 4 −1
0 −1 0
(i) [3 pts] Apply the Laplacian mask at position (2,2) (0-based indexing,
center at row 2, column 2). Write the full expression (sum of products).
(ii) [3 pts] Apply the same mask at position (1,1) (top-left corner of the inner
3×3 region). Show all working.
(iii) [2 pts] Compare the two results. What does the sign (positive/negative)
of the Laplacian response indicate about the local image structure?
Problem 2 [15 pts]
Two sets of 2D points are related by an affine transformation (6 degrees
of freedom):
Roll number:
Name:
{
'
x =ax +by +t x
y ' =cx +dy +t y
(a) [3 pts] Write this transformation as a 3×3 homogeneous matrix.
You are given the following point correspondences (with possible noise):
Point (x , y) ( x' , y' )
1 (0, 0) (2, 3)
2 (1, 0) (4, 3)
3 (0, 1) (1, 5)
4 (1, 1) (3, 5)
5 (2, 3) (5, 9)
(b) [5 pts] How would you solve for the 6 unknown parameters using Least
Squares? Write down the exact matrices A and b in the system A x=b ,
where x=¿ .
(c) [3 pts] What is the minimum number of point
correspondences required to estimate this transformation? Justify.
(d) [4 pts] Describe how RANSAC would be used to robustly estimate the
affine transformation when outliers are present. Include the steps of
hypothesis generation and consensus set evaluation.
Problem 3 [8 pts]
Match each frequency-domain operation (left) with its spatial-domain effect
(right). Write the letter next to the number.
Frequency Domain Operation Spatial Domain Effect
1. Multiply Fourier transform by a A. The image is convolved with a Gaussian kernel; no ringing.
binary mask that is 1 inside a
low-frequency circle, 0 elsewhere.
2. Multiply Fourier transform by a B. The image is convolved with a sinc-like kernel; ringing
binary mask that is 0 inside a narrow artifacts appear.
low-frequency disc, 1 elsewhere.
3. Multiply Fourier transform by a C. The image is blurred (low-pass filtered).
Gaussian function centered at the
origin.
4. Set a specific frequency ring (e.g., D. A periodic texture (e.g., a grid pattern) is removed.
radius 20–25) to zero.
Roll number:
Name:
Problem 4 [8 pts]
You are given a small 3×3 image patch of grayscale intensities:
[ ]
50 70 80
I = 60 90 100
55 85 95
The Harris corner detector computes the matrix
[ ]
2
∑ Ix ∑ Ix I y
M= 2
∑I x I y ∑Iy
over the local window (here, the entire 3×3 patch).
Use central differences for derivatives:
I x (i , j)=I (i , j+1)−I (i, j−1) (with zero outside boundaries)
I y (i, j)=I (i+1 , j)−I (i−1, j) (with zero outside boundaries).
(i) [3 pts] Compute I x and I y for each pixel in the 3×3 patch (write the 3×3
results).
(ii) [3 pts] Compute the 2 ×2 matrix M for this patch (sum over all 9 pixels).
Show your work.
(iii) [2 pts] Compute the two eigenvalues λ 1 and λ 2 of M (you may leave the
square root expression unsimplified if needed). Based on the eigenvalues,
classify this patch as flat, edge, or corner. Briefly explain.
Problem 5 [6 pts]
For each 3×3 homogeneous transformation matrix, describe in words what geometric change it
applies to a point ( x , y ) in the image plane.
(i)
[ ]
1 0 −20
H 1= 0 1 15
0 0 1
(ii)
[ ]
cos 45 ∘ −sin 45 ∘ 0
H 2= sin 45∘ cos 45∘ 0
0 0 1
(iii)
Roll number:
Name:
[ ]
1.5 0.3 0
H 3= 0.2 1.2 0
0 0 1
(iv)
[ ]
1 0 0
H4= 0 1 0
0.002 −0.001 1
Problem 6 [8 pts]
Answer each in 2–3 sentences.
(i) You have a single image of a flat rectangular table. You know the actual
width of the table. Explain how you could compute the actual length of the
table from the image using a homography. What assumption must you make
about the camera?
(ii) Describe how placing a reference object of known length (e.g., a 1 m
ruler) on the same plane allows you to measure other distances on that
plane without explicitly computing the homography. What condition must
hold for this to work?