The IIT Madras BS Deep Learning for Computer Vision (Deep Learning for Computer Vision) End Term paper sat on 10 May 2026, in the January 2026 term: 22 questions for 57 marks in 180 minutes. Every question is below with its answer. Take it as a timed mock test to be marked, or read it through first.
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)?
256×256 = 65,536 bytes
256×256×2 = 131,072 bytes
256×256×3 = 196,608 bytes
256×256×3×8 = 1,572,864 bytes
Correct answer
256×256×3 = 196,608 bytes
Which statement is correct?
Convolution equals correlation for all filters.
Convolution flips the kernel spatially; correlation does not.
Correlation flips the kernel; convolution does not.
Both operations always preserve energy.
Correct answer
Convolution flips the kernel spatially; correlation does not.
You apply a 5×5 filter with stride 1 and no padding on a 32×32 image. Output size is:
32×32
28×28
27×27
26×26
Correct answer
28×28
The Sobel operator primarily estimates:
Correct answer
SIFT is designed to be robust mainly to:
Scale and rotation changes
Only translation (shifts)
Only illumination changes
Only affine viewpoint changes with no residual error
Correct answer
Scale and rotation changes
In classical scale-space theory, increasing σ in Gaussian smoothing generally:
Enhances high-frequency details
Suppresses high-frequency details and removes fine structures
Has no effect on edges
Increases noise
Correct answer
Suppresses high-frequency details and removes fine structures
Backpropagation is best described as:
A method to avoid overfitting
Efficient computation of gradients using the chain rule through computation graphs
A method that works only for linear models
An optimizer like Adam
Correct answer
Efficient computation of gradients using the chain rule through computation graphs
Which is a key effect of momentum?
Makes gradients unbiased
Accumulates a velocity term
Gurantees to converge unlike vanilla Gradient Descent
Eliminates need for setting learning rate by adpatly calculaitng it on the fly.
Correct answer
Accumulates a velocity term
L2 regularization on weights most directly encourages:
Sparse weights with many exact zeros
Smaller-magnitude weights
Larger-magnitude weights
Larger learning rates
Correct answer
Smaller-magnitude weights
A key reason CNNs are parameter-efficient compared to fully connected nets for images is:
Weight sharing and local receptive fields
They never use nonlinearities thus easier to train
They always use pooling
Compared to vanilla neural network and RNN, they are robust to gradient explosion and vanishing gradient.
Correct answer
Weight sharing and local receptive fields
Stacking multiple 3×3 convolutions (stride 1) increases receptive field because:
Each layer composes local neighborhoods into larger effective context
Each layer reduces spatial resolution
It removes nonlinearities thereby increasing the effective context
It forces global averaging thereby increasing the contribution of input pixels
Correct answer
Each layer composes local neighborhoods into larger effective context
In Faster R-CNN, the Region Proposal Network (RPN) outputs:
Only class probabilities + pixel-wise masks
Objectness scores + bounding box proposals
Pixel-wise masks
Optical flow
Correct answer
Objectness scores + bounding box proposals
Which statement is correct ?
Semantic segmentation separates individual object instances.
Instance segmentation distinguishes different object instances of the same class.
Both are identical tasks.
Neither predicts labels per pixel.
Correct answer
Instance segmentation distinguishes different object instances of the same class.
Vanishing gradients in vanilla RNNs occur mainly due to:
High deapth of the RNN layers
Repeated multiplication by Jacobians across timesteps
Too many convolution layers before RNN
Use of softmax
Correct answer
Repeated multiplication by Jacobians across timesteps
Select all correct statements:
Harris corner detector responds strongly where gradient changes in two orthogonal directions.
Difference-of-Gaussians approximates Laplacian-of-Gaussian for blob detection.
SIFT descriptors are computed from raw pixel intensities without gradients.
Canny edge detector is a data driven method to calculate edges in an image.
Correct answers
Harris corner detector responds strongly where gradient changes in two orthogonal directions.
Difference-of-Gaussians approximates Laplacian-of-Gaussian for blob detection.
Select all correct statements :
Adam adapts per-parameter learning rates using estimates of first and second moments.
Early stopping can act as a form of regularization.
Dropout increases training accuracy by removing bad training data samples .
Data augmentation typically improves generalization in vision tasks.
Correct answers
Adam adapts per-parameter learning rates using estimates of first and second moments.
Early stopping can act as a form of regularization.
Data augmentation typically improves generalization in vision tasks.
Select all correct statements:
Self-attention computes pairwise interactions between tokens within a sequence.
Positional encodings are needed because self-attention alone is permutation- invariant.
Transformers must use recurrence to handle sequences.
Multi-head attention allows the model to attend to different subspaces/relations.
Correct answers
Self-attention computes pairwise interactions between tokens within a sequence.
Positional encodings are needed because self-attention alone is permutation- invariant.
Multi-head attention allows the model to attend to different subspaces/relations.
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.
Correct answer: 0.143 (accepted within ±0.001)
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.
Correct answer: 0.41 (accepted within ±0.002)
Input image: 64×64. Convolution: kernel 7×7, stride 2, padding 3. Compute output spatial size.
Correct answer: 32
A fully connected network maps input 100 → hidden 50 → output 10. Ignore bias. What is the total number of weights?
Correct answer: 5500
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?
Correct answer: 8768
