AVD624/ESG667 Pre-Mid Exam February 2026
Indian Institute of Space Science and Technology
Department of Avionics
AVD624/ESG667 Computer Vision Pre-Mid Exam
Total Marks: 20 Duration: 40 Minutes
Questions: 4 Marks per Question: 5
Instructions: Answer all questions. No extra sheets will be provided. Marks for each sub-part
are indicated on the right. All the best!
Question 1. Homography Estimation via Direct Linear Transform (DLT)
Consider the following four point correspondences between two planes:
x1 = (0, 0)⊤ ↔ x′1 = (1, 1)⊤ , x2 = (1, 0)⊤ ↔ x′2 = (4, 1)⊤ ,
x3 = (1, 1)⊤ ↔ x′3 = (4, 4)⊤ , x4 = (0, 1)⊤ ↔ x′4 = (1, 4)⊤ .
(a) For a single correspondence (x, y) ↔ (x′ , y ′ ), the DLT formulation gives a 2 × 9 matrix Ai
such that Ai h = 0, where h = vec(H) is the vectorisation of the homography. Write the
explicit Ai for the first correspondence x1 ↔ x′1 . [1 marks]
(b) State the minimum number of correspondences required to solve for H via DLT. Specify
the geometric constraint that the chosen points must satisfy. [1 mark]
(c) After stacking all Ai into matrix A and solving Ah = 0 via SVD, which singular vector
yields ĥ? Justify why that choice minimises ∥Ah∥ subject to ∥h∥ = 1. [3 marks]
Question 2. Camera Projection and Zhang’s Calibration Method
A calibrated camera has the intrinsic matrix
500 0 320
K = 0 500 240 .
0 0 1
(a) A 3-D world point Xw = (1, 0, 5)⊤ (metres) is projected with R = I and t = 0. Write the
full projection equation λ x̃ = K [R | t] X̃w and compute the pixel coordinates (u, v). [2
marks]
(b) The camera is now translated to t = (−1, 0, 0)⊤ with R = I unchanged. Recompute (u, v).
Using both results, explain how lateral translation affects pixel displacement and how the
focal length f mediates this relationship. [2 marks]
(c) In Zhang’s method the calibration target is a planar checkerboard with Z = 0. Write the
simplified projection matrix under this assumption, and state in one sentence why this
planar target is sufficient to recover all five intrinsic parameters. [1 mark]
Page 1 of 2 Total: 20 Marks
AVD624/ESG667 Pre-Mid Exam February 2026
Question 3. Linear Shift-Invariant Systems and Image Filtering
Consider the 5 × 5 image patch f [n, m] and 3 × 3 box filter h[n, m] below (top-left pixel is index
[0, 0]):
0 0 0 0 0
0 90 90 90 0 1 1 1
1
f =0 90 0 90 0, h=
9
1 1 1 .
0 90 90 90 0 1 1 1
0 0 0 0 0
(a) Compute the 2-D convolution output g[2, 2] (the centre pixel). [1.5 marks]
(b) Determine whether the moving-average filter h is an LTI system by explicitly verifying
both required properties:
(i) superposition and (ii) time-invariance. [2 marks]
(c) Give two specific advantages of a Gaussian filter over the uniform box filter h for image
smoothing. [1.5 marks]
Question 4. Edge Detection: Image Gradients and the Canny Detector
(a) The Canny pipeline includes a step called Non-Maximum Suppression (NMS).
(i) Describe precisely what NMS does and why it is necessary after computing gradient
magnitudes. [1 mark]
(ii) A pixel at (x, y) has gradient magnitude |∇G| = 80 and gradient direction θ = 90◦ .
Its two neighbours along the gradient direction have magnitudes 70 and 85. Does this
pixel survive NMS? [0.5 marks]
(b) Canny uses hysteresis thresholding with TH = 80 and TL = 40. Classify each pixel and
state the final edge decision for Pixels B and C:
Pixel Gradient Magnitude Connectivity
A 120 —
B 55 Connected to Pixel A
C 55 Not connected to any strong edge pixel
[1.5 marks]
(c) The Sobel operator approximates the image gradient using two 3 × 3 kernels.
(i) Write the standard Sobel kernels Gx and Gy . [1 mark]
(ii) Given gradient components Gx = 3 and Gy = 4 at a pixel, compute the gradient
magnitude |∇I| and the gradient direction θ (in degrees). [1 mark]
— End of Question Paper —
Page 2 of 2 Total: 20 Marks