Home / Study library / Deep Learning for Computer Vision / End Term · Jan 2026 Deep Learning for Computer Vision · End Term · Jan 2026 · FN
Deep Learning for Computer Vision · End Term · Jan 2026 FN
← Course papers · Start practice / exam
Questions and published explanations below are available without starting a test. Some questions may not have a published solution yet.
0 / 22 published solutions
Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10 Q11 Q12 Q13 Q14 Q15 Q16 Q17 Q18 Q19 Q20 Q21 Q22 Q23 Question 2 MCQ · 2.0 marks A standard RGB image of size 256×256 is stored as 8-bit unsigned integers per channel. What is
the total memory (in bytes) to store the raw image (no compression)?
6406536959477. 256×256 = 65,536 bytes
6406536959478. 256×256×2 = 131,072 bytes
6406536959479. 256×256×3 = 196,608 bytes
6406536959480. 256×256×3×8 = 1,572,864 bytes
A published solution is not available for this question yet.
Report a correction for Q2
Question 3 MCQ · 2.0 marks Which statement is correct?
6406536959481. Convolution equals correlation for all filters.
6406536959482. Convolution flips the kernel spatially; correlation does not.
6406536959483. Correlation flips the kernel; convolution does not.
6406536959484. Both operations always preserve energy.
A published solution is not available for this question yet.
Report a correction for Q3
Question 4 MCQ · 2.0 marks You apply a 5×5 filter with stride 1 and **no padding** on a 32×32 image. Output size is:
A published solution is not available for this question yet.
Report a correction for Q4
Question 5 MCQ · 2.0 marks The Sobel operator primarily estimates:
6406536959489. 6406536959490. 6406536959491. 6406536959492. A published solution is not available for this question yet.
Report a correction for Q5
Question 6 MCQ · 2.0 marks SIFT is designed to be robust mainly to:
6406536959493. Scale and rotation changes
6406536959494. Only translation (shifts)
6406536959495. Only illumination changes
6406536959496. Only affine viewpoint changes with no residual error
A published solution is not available for this question yet.
Report a correction for Q6
Question 7 MCQ · 2.0 marks In classical scale-space theory, increasing σ in Gaussian smoothing generally:
6406536959497. Enhances high-frequency details
6406536959498. Suppresses high-frequency details and removes fine structures
6406536959499. Has no effect on edges
6406536959500. Increases noise
A published solution is not available for this question yet.
Report a correction for Q7
Question 8 MCQ · 2.0 marks Backpropagation is best described as:
6406536959501. A method to avoid overfitting
6406536959502. Efficient computation of gradients using the chain rule through computation
graphs
6406536959503. A method that works only for linear models
6406536959504. An optimizer like Adam
A published solution is not available for this question yet.
Report a correction for Q8
Question 9 MCQ · 2.0 marks Which is a key effect of momentum?
6406536959505. Makes gradients unbiased
6406536959506. Accumulates a velocity term
6406536959507. Gurantees to converge unlike vanilla Gradient Descent
6406536959508. Eliminates need for setting learning rate by adpatly calculaitng it on the fly.
A published solution is not available for this question yet.
Report a correction for Q9
Question 10 MCQ · 2.0 marks L2 regularization on weights most directly encourages:
6406536959509. Sparse weights with many exact zeros
6406536959510. Smaller-magnitude weights
6406536959511. Larger-magnitude weights
6406536959512. Larger learning rates
A published solution is not available for this question yet.
Report a correction for Q10
Question 11 MCQ · 2.0 marks A key reason CNNs are parameter-efficient compared to fully connected nets for images is:
6406536959513. Weight sharing and local receptive fields
6406536959514. They never use nonlinearities thus easier to train
6406536959515. They always use pooling
6406536959516. Compared to vanilla neural network and RNN, they are robust to gradient
explosion and vanishing gradient.
A published solution is not available for this question yet.
Report a correction for Q11
Question 12 MCQ · 2.0 marks Stacking multiple 3×3 convolutions (stride 1) increases receptive field because:
6406536959517. Each layer composes local neighborhoods into larger effective context
6406536959518. Each layer reduces spatial resolution
6406536959519. It removes nonlinearities thereby increasing the effective context
6406536959520. It forces global averaging thereby increasing the contribution of input pixels
A published solution is not available for this question yet.
Report a correction for Q12
Question 13 MCQ · 2.0 marks In Faster R-CNN, the Region Proposal Network (RPN) outputs:
6406536959521. Only class probabilities + pixel-wise masks
6406536959522. Objectness scores + bounding box proposals
6406536959523. Pixel-wise masks
6406536959524. Optical flow
A published solution is not available for this question yet.
Report a correction for Q13
Question 14 MCQ · 2.0 marks Which statement is correct ?
6406536959525. Semantic segmentation separates individual object instances.
6406536959526. Instance segmentation distinguishes different object instances of the same
class.
6406536959527. Both are identical tasks.
6406536959528. Neither predicts labels per pixel.
A published solution is not available for this question yet.
Report a correction for Q14
Question 15 MCQ · 2.0 marks Vanishing gradients in vanilla RNNs occur mainly due to:
6406536959529. High deapth of the RNN layers
6406536959530. Repeated multiplication by Jacobians across timesteps
6406536959531. Too many convolution layers before RNN
6406536959532. Use of softmax
A published solution is not available for this question yet.
Report a correction for Q15
Question 16 MSQ · 3.0 marks Select all correct statements:
6406536959533. Harris corner detector responds strongly where gradient changes in two
orthogonal directions.
6406536959534. Difference-of-Gaussians approximates Laplacian-of-Gaussian for blob
detection.
6406536959535. SIFT descriptors are computed from raw pixel intensities without gradients.
6406536959536. Canny edge detector is a data driven method to calculate edges in an image.
A published solution is not available for this question yet.
Report a correction for Q16
Question 17 MSQ · 3.0 marks Select all correct statements :
6406536959537. Adam adapts per-parameter learning rates using estimates of first and second
moments.
6406536959538. Early stopping can act as a form of regularization.
6406536959539. Dropout increases training accuracy by removing bad training data samples .
6406536959540. Data augmentation typically improves generalization in vision tasks.
A published solution is not available for this question yet.
Report a correction for Q17
Question 18 MSQ · 3.0 marks Select all correct statements:
6406536959541. Self-attention computes pairwise interactions between tokens within a
sequence.
6406536959542. Positional encodings are needed because self-attention alone is permutation-
invariant.
6406536959543. Transformers must use recurrence to handle sequences.
6406536959544. Multi-head attention allows the model to attend to different
subspaces/relations.
A published solution is not available for this question yet.
Report a correction for Q18
Question 19 NAT · 4.0 marks Box A: top-left (0,0), bottom-right (8,8).
Box B: top-left (4,4), bottom-right (12,12).
Compute IoU as a decimal.
A published solution is not available for this question yet.
Report a correction for Q19
Question 20 NAT · 4.0 marks For a single query, suppose similarity with the positive is s\(^{+}\)=2 and with two negatives are s1=1,
s2=0.
InfoNCE loss: L = −log( exp(s\(^{+}\)) / (exp(s\(^{+}\))+exp(s1)+exp(s2)) ).
Using exp(2)=7.39, exp(1)=2.72, exp(0)=1, compute L.
A published solution is not available for this question yet.
Report a correction for Q20
Question 21 NAT · 4.0 marks Input image: 64×64. Convolution: kernel 7×7, stride 2, padding 3. Compute output spatial size.
A published solution is not available for this question yet.
Report a correction for Q21
Question 22 NAT · 4.0 marks A fully connected network maps input 100 → hidden 50 → output 10. Ignore bias. What is the total
number of weights?
A published solution is not available for this question yet.
Report a correction for Q22
Question 23 NAT · 4.0 marks Input feature map has M=64 channels. You apply depthwise 3×3 followed by pointwise 1×1 to
produce N=128 channels. Ignore bias. What is the Total parameters?
A published solution is not available for this question yet.
Report a correction for Q23
