Final Viva: Machine Vision & Motion Analysis

On the five questions

Spring 2026

Section A
Question 1
(a) Analyze why an image with a high variance (σ 2 ) generally exhibits higher contrast than
one with low variance, referencing the mathematical definition of contrast.
Solution: (Reference: mv&ma_01_problems.pdf,
PL−1 Question 1)
2 2
Variance is defined as σ = k=0 (rk − µ) pr (rk ). High variance indicates that pixel intensities
are spread far from the mean, covering a wider portion of the dynamic range [0, L − 1], which is
the primary goal of contrast enhancement.

(b) Explain why a Median Filter is preferred over an Averaging (Box) Filter when dealing
specifically with ”salt-and-pepper” noise.
Solution: (Reference: mv&ma_01_problems.pdf, Question 2)
A Box filter blurs edges and is ineffective against impulse noise because it averages the outlier
(noise) into the neighborhood. The Median filter sorts neighborhood pixels and selects the
median value, making it robust to outliers and highly effective at removing impulse noise while
preserving edges.

Question 2
(a) Evaluate the visual phenomenon observed when recording a car tire rotating at a fre-
quency of 60 Hz using a standard smartphone camera capturing at 30 frames per second
(FPS). Propose a mathematical justification for the resulting visual artifact based on tem-
poral sampling principles.
Solution: (Reference: mv&ma_02_problems.pdf, Problem 1)
The visual artifact observed is temporal aliasing, commonly known as the ”wagon-wheel effect,”
where the tire may appear to rotate backward, slowly, or remain stationary. Mathematically,
this violates the Nyquist-Shannon sampling theorem, which dictates that the temporal sampling
rate (fs ) must be strictly greater than twice the maximum motion frequency (fmax ) to accurately
reconstruct motion: fs > 2fmax . Here, fs = 30 Hz and 2fmax = 120 Hz. Because 30 < 120, the
motion is undersampled, causing consecutive frames to capture the spokes in near-identical
but slightly delayed positions, creating the illusion of backward or distorted motion.

(b) Analyze the impact of losing an I-frame versus a B-frame during a live Zoom lecture
over an unstable Wi-Fi network. Deduce the visual consequences on the receiver’s screen
based on the structural hierarchy of a Group of Pictures (GOP).

1
Solution: (Reference: mv&ma_02_problems.pdf, Problem 2)
Based on the GOP structure, I-frames (Intra-coded) are fully independent frames compressed
using spatial redundancy, acting as the anchor for the sequence. Losing an I-frame causes se-
vere, prolonged visual artifacts (heavy pixelation or screen tearing) that persist until the next
I-frame arrives, because all subsequent P and B frames rely on it for temporal prediction. Con-
versely, B-frames (Bi-directionally predicted) rely on past and future frames to encode only dif-
ferences. Losing a B-frame results in only a momentary glitch or micro-stutter for that specific
fraction of a second, as no other frames depend on it for their decoding.

Question 3
(a) Evaluate the robustness of the Brightness Constancy Assumption for a stationary surveil-
lance camera monitoring a parking lot as dense clouds rapidly pass over the sun. Deter-
mine the mathematical consequence on the Optical Flow Constraint Equation (OFCE).
Solution: (Reference: mv&ma_03_problems.pdf, Question 1)
The OFCE assumes I(x, y, t) = I(x+u∆t, y +v∆t, t+∆t), meaning physical objects do not change
brightness as they move. A passing shadow rapidly changes the temporal intensity (It 6= 0)
while the physical velocity of the parked cars is zero. The OFCE will erroneously compute a non-
zero flow (u, v) to mathematically compensate for the intensity change, resulting in ”phantom”
motion fields where no physical motion occurred.

(b) Consider a smartphone camera closely scanning a horizontal barcode moving verti-
cally across the screen. Analyze the fundamental limitation a local sparse optical flow
algorithm faces in this scenario based on the image gradient.
Solution: (Reference: mv&ma_03_problems.pdf, Question 2)
This scenario demonstrates the Aperture Problem. Because the barcode consists solely of hor-
izontal edges, the vertical gradient Iy is strong, but the horizontal gradient Ix is zero. The con-
straint becomes a single line in (u, v) space. The system is locally under-determined in the direc-
tion parallel to the edge, making it impossible to calculate the true 2D velocity vector without
observing corners or incorporating a global smoothness constraint.

Section B
Question 4
(a) Stereo Vision & Depth Estimation: Given a stereo camera setup with a baseline b = 0.6
m and focal length f = 1000 pixels, calculate the exact depth Z for a corresponding point
with a disparity of d = 30 pixels. Determine the mathematical sensitivity (derivative) of
the depth with respect to a 1-pixel measurement error in disparity.
Solution: (Reference: mv&ma_04_problems.pdf, Solution 1)
The formula for depth is Z = fd·b .
Given f = 1000, b = 0.6 and d = 30, Z = 1000·0.6
30 = 600
30 = 20 meters.
To find the sensitivity to disparity error, compute the derivative of Z with respect to d:
dZ f ·b
dd = − d2
For d = 30: dZ 600 600
dd = − 302 = − 900 ≈ −0.67 meters/pixel.
A 1-pixel error in disparity alters the depth estimation by approximately 0.67 meters.

2
(b) Lucas-Kanade Optical Flow: Formulate the matrix equation for the Lucas-Kanade op-
tical flow constraint. State the algebraic condition required for the 2 × 2 structural tensor
matrix to be invertible, and relate this condition to the visual texture of the tracked image
patch.
Solution: (Reference: mv&ma_04_problems.pdf, Solution 2)
The least-squares matrix equation is:
 P 2 P   P 
Ix Ix I y u Ix I t
P P 2 =− P
Ix Iy Iy v Iy It

For this system to be solvable, the 2 × 2 matrix T

P A2 AP(the
2
structural
P 2
tensor) must be invertible,
meaning its determinant must be non-zero: ( Ix )( Iy ) − ( Ix Iy ) > 0. Visually, this requires
the image patch to have gradients in at least two different directions (i.e., corners or highly
textured regions). A flat region or a straight edge results in a singular matrix (the aperture
problem).

Question 5
(a) Pipeline Justification: Given a high-noise manufacturing environment capturing im-
ages of stamped metal parts, justify the selection of the Canny edge detector over a stan-
dard 3 × 3 Sobel operator. Explicitly reference the internal mathematical stages of the
Canny pipeline that manage this noise constraint.
Solution: (Reference: mv&ma_05_problems.pdf, Sol 1)
The standard Sobel operator merely applies a first-order derivative filter. While it has mild
smoothing (weight of 2 in the center), raw gradients strictly amplify noise, leading to thick,
isotropic, and noisy edges. Canny is justified because its pipeline explicitly combats noise: 1) It
begins with a strict Gaussian Smoothing pass. 2) It uses Non-Maxima Suppression to thin the
gradient responses down to a precise 1-pixel continuous edge. 3) It applies Hysteresis Thresh-
olding (dual thresholds Th and Tl ) to rescue weak edge signals only if they are geometrically
connected to strong edges, filtering out isolated high-gradient noise artifacts.

(b) Computational Optimization: Evaluate the necessity of calculating the response

score R = det(M ) − k[tr(M )]2 in the Harris Corner Detector instead of directly computing
the eigenvalues (λ1 , λ2 ) of the structure tensor M for a real-time mobile augmented reality
application.
Solution: (Reference: mv&ma_05_problems.pdf, Sol 2)
Extracting eigenvalues λ1 and λ2 requires solving the characteristic polynomial of the 2×2 struc-
ture tensor M , which involves computationally expensive floating-point square root operations
for every single pixel. In a real-time mobile AR application, compute budgets are strict. The
response function R = det(M ) − k[tr(M )]2 cleverly uses the determinant (λ1 λ2 ) and the trace
(λ1 + λ2 ). This requires only basic multiplication and addition of the matrix elements, entirely
bypassing the square root bottleneck while still mathematically identifying regions where both
eigenvalues are large (corners).