Figure from the original question paper

(B) Option figure from the original question paper

(A) Option figure from the original question paper

(C) Option figure from the original question paper

(D) Option figure from the original question paper

Match the following: | 1) VGGNet | i) 1 \times 1 convolution | |---|---| | 2) EfficientNet | ii) identity mapping | | 3) GoogleNet | iii) 3 \times 3 convolution | | 4) ResNet | iv) 7 \times 7 convolution | | | v) depth-wise separable convolutions |

(D) 1→iii, 2→v, 3→i, 4→ii

(A) 1→iv, 2→ii, 3→v, 4→i

(B) 1→iv, 2→iii, 3→v, 4→ii

(C) 1→iv, 2→v, 3→ii, 4→iii

Figure from the original question paper

(B) Softmax of alignment scores

(A) Sigmoid of alignment scores

Match the following computer vision tasks to situations: | 1) Instance Segmentation | i) There are cars in these pixels | |---|---| | 2) Classification | ii) There are 4 cars in the image | | 3) Semantic Segmentation | iii) There is car in the image | | 4) Object Detection | iv) There are 4 cars in the image, and here are the pixels with the shape of each of their occurrence |

(C) 1→ iv, 2→ iii, 3→ i, 4 → ii

(A) 1→ iv, 2→ iii, 3→ ii, 4 → i

(B) 1→ ii, 2→ iii, 3 → i, 4 →iv

(D) 1→ ii, 2→ iv, 3 → iii, 4→ i

Match the following: | 1) One-to-many RNN architecture | i) Sentiment analysis of Netflix reviews | |---|---| | 2) Many-to-one RNN architecture | ii) Convert English sentence to Hindi | | 3) Many-to-many RNN architecture with equal number of inputs and outputs | iii) Named entity recognition | | 4) Many-to-many RNN architecture with unequal number of inputs and outputs (Encoder and Decoder model) | iv) Music generation |

(B) 1→ iv, 2→ i, 3 → iii, 4→ ii

(A) 1→ iii, 2→ iv, 3→ ii, 4 → i

(C) 1→ i, 2→ ii, 3 → iii, 4→ iv

(D) 1→ i, 2→ iii, 3 → ii, 4→ iv

Match the following attention mechanisms to their corresponding alignment score functions: | (1) General Attention | (i) v_a^\top \tanh(W_a[s_t; h_i]) | |---|---| | (2) Content-Based Attention | (ii) \alpha_{t,j} = \text{softmax}(W_a s_t) | | (3) Dot-Product Attention | (iii)s_t^\top h_i | | (4) Additive Attention | (iv) \cos(s_t, h_i) | | (5) Location-Based Attention | (v) s_t^\top W_a h_i |

(D) 1→v, 2→iv, 3→iii, 4→i,5→ii

(A) 1→iii, 2→ii, 3→v, 4→iv,5→i

(B) 1→iii, 2→iv, 3→i, 4→ii,5→v

(C) 1→i, 2→iii, 3→iv, 4→ii,5→v

January 2025 term · Deep Learning for Computer Vision · BSDA5006

Deep Learning for Computer Vision Quiz 2: 16 March 2025 (January 2025 term)

The IIT Madras BS Deep Learning for Computer Vision (Deep Learning for Computer Vision) Quiz 2 paper sat on 16 Mar 2025, in the January 2025 term: 37 questions for 65 marks in 120 minutes. Every question is below with its answer. Take it as a timed mock test to be marked, or read it through first.

Questions: 37
Marks: 65
Duration: 120 min
MCQ: 16
MSQ: 3
Numerical: 18

Take as mock test Practise with answers

Updated 25 Sept 2026

Official paper: IIT M IMPROVEMENT AN EXAM QIM2 16 Mar 2025 · No negative marking.

Question 1

+2 marksOne correct option

Show answer

Correct answer

Question 1 on its own page

Question 2

+2 marksOne correct option

Which one of the following object detection networks uses an ROI pooling layer?

A
Fast R-CNN
B
R-CNN
C
YOLO
D
All of these

Show answer

Correct answer

A
Fast R-CNN

Question 2 on its own pageAlso asked in Quiz 2 1 Dec 2024

Question 3

+2 marksOne correct option

Match the following:

1) VGGNet	i) $1 \times 1$ convolution
2) EfficientNet	ii) identity mapping
3) GoogleNet	iii) $3 \times 3$ convolution
4) ResNet	iv) $7 \times 7$ convolution
	v) depth-wise separable convolutions

A
1→iv, 2→ii, 3→v, 4→i
B
1→iv, 2→iii, 3→v, 4→ii
C
1→iv, 2→v, 3→ii, 4→iii
D
1→iii, 2→v, 3→i, 4→ii

Show answer

Correct answer

D
1→iii, 2→v, 3→i, 4→ii

Question 3 on its own page

Question 4

+2 marksOne correct option

Which one of the following statements is False?

A
In a max pooling layer, the unit that gets selected (maximum entry) in forward propagation gets all the gradient during backpropagation.
B
Residual blocks are used in ResNets to avoid the vanishing gradient problem.
C
Feature maps of initial layers of a CNN capture more abstract concepts than the feature maps of final layers.
D
Dilated convolution increases the receptive field size when compared to the standard convolution operator.

Show answer

Correct answer

C
Feature maps of initial layers of a CNN capture more abstract concepts than the feature maps of final layers.

Question 4 on its own pageAlso asked in Quiz 2 4 Aug 2024, Quiz 2 1 Dec 2024

Question 5

+2 marksOne correct option

A
Sigmoid of alignment scores
B
Softmax of alignment scores
C
ReLU of alignment scores
D
None of these

Show answer

Correct answer

B
Softmax of alignment scores

Question 5 on its own page

Question 6

+2 marksOne correct option

In object detection using CNNs, which loss function is typically used to optimize the confidence scores assigned to predicted bounding boxes, aiming to ensure high confidence for correct detec- tions and penalize false positives?

A
Focal Loss
B
Smooth Ll Loss
C
Binary Cross-Entropy Loss
D
Mean Squared Error (MSE)

Show answer

Correct answer

C
Binary Cross-Entropy Loss

Question 6 on its own page

Question 7

+2 marksOne correct option

Consider two 12 × 12 bounding boxes(one on the upper left and one of the lower right) in an images with an overlapping region of 8 × 8. The Intersection over Union (IoU) between the two boxes is (choose the closest value):

A
21%
B
25%
C
28%
D
37%

Show answer

Correct answer

C
28%

Question 7 on its own pageAlso asked in Quiz 2 1 Dec 2024

Question 8

+2 marksOne correct option

In object detection using CNNs, what is the challenge addressed by the non-maximum suppression (NMS) technique?

A
Handling variations in object scale and aspect ratio
B
Improving the localization accuracy of detected objects
C
Reducing the computational complexity of the detection process
D
Removing redundant bounding box predictions for the same object

Show answer

Correct answer

D
Removing redundant bounding box predictions for the same object

Question 8 on its own page

Question 9

+2 marksOne correct option

Match the following computer vision tasks to situations:

1) Instance Segmentation	i) There are cars in these pixels
2) Classification	ii) There are 4 cars in the image
3) Semantic Segmentation	iii) There is car in the image
4) Object Detection	iv) There are 4 cars in the image, and here are the pixels with the shape of each of their occurrence

A
1→ iv, 2→ iii, 3→ ii, 4 → i
B
1→ ii, 2→ iii, 3 → i, 4 →iv
C
1→ iv, 2→ iii, 3→ i, 4 → ii
D
1→ ii, 2→ iv, 3 → iii, 4→ i

Show answer

Correct answer

C
1→ iv, 2→ iii, 3→ i, 4 → ii

Question 9 on its own pageAlso asked in Quiz 2 23 Nov 2025

Question 10

+2 marksOne correct option

Match the following:

1) One-to-many RNN architecture	i) Sentiment analysis of Netflix reviews
2) Many-to-one RNN architecture	ii) Convert English sentence to Hindi
3) Many-to-many RNN architecture with equal number of inputs and outputs	iii) Named entity recognition
4) Many-to-many RNN architecture with unequal number of inputs and outputs (Encoder and Decoder model)	iv) Music generation

A
1→ iii, 2→ iv, 3→ ii, 4 → i
B
1→ iv, 2→ i, 3 → iii, 4→ ii
C
1→ i, 2→ ii, 3 → iii, 4→ iv
D
1→ i, 2→ iii, 3 → ii, 4→ iv

Show answer

Correct answer

B
1→ iv, 2→ i, 3 → iii, 4→ ii

Question 10 on its own pageAlso asked in Quiz 2 1 Dec 2024, Quiz 2 23 Nov 2025

Question 11

+2 marksOne correct option

Match the following attention mechanisms to their corresponding alignment score functions:

(1) General Attention	(i) $v_a^\top \tanh(W_a[s_t; h_i])$
(2) Content-Based Attention	(ii) $\alpha_{t,j} = \text{softmax}(W_a s_t)$
(3) Dot-Product Attention	(iii) $s_t^\top h_i$
(4) Additive Attention	(iv) $\cos(s_t, h_i)$
(5) Location-Based Attention	(v) $s_t^\top W_a h_i$

A
1→iii, 2→ii, 3→v, 4→iv,5→i
B
1→iii, 2→iv, 3→i, 4→ii,5→v
C
1→i, 2→iii, 3→iv, 4→ii,5→v
D
1→v, 2→iv, 3→iii, 4→i,5→ii

Show answer

Correct answer

D
1→v, 2→iv, 3→iii, 4→i,5→ii

Question 11 on its own page

Question 12

+2 marksOne correct option

Which of the following techniques help control the exploding or vanishing gradient problem in recurrent neural networks?
1. Gradient clipping
2. Use ReLU activation funtion to control exploding gradient and sigmoid activation to control vanishing gradients
3. Use Long Short-Term Memory (LSTM)
4. Use data augmentation methods (for example, reversing the input sequence)

A
1, 2, and 3; but not 4
B
1, 3, and 4; but not 2
C
All of these
D
Only 1 and 3

Show answer

Correct answer

B
1, 3, and 4; but not 2

Question 12 on its own page

Question 13

+2 marksOne correct option

State whether the given statement is True or False:
While training an object detection algorithm, you need a training set that contains many pictures of the objects you want to detect. However, bounding boxes are not typically provided in the training set, since architectures like Faster-RCNN and YOLO can learn to detect the object by themselves.

A
True
B
False

Show answer

Correct answer

B
False

Question 13 on its own page

Question 14

+2 marksOne correct option

How are saliency maps generated in the context of Convolutional Neural Networks (CNNs)?

A
By applying a convolutional filter to the input image
B
By computing the gradient of the model’s output with respect to the input image
C
By randomly sampling regions of the image and assessing their importance
D
By measuring the similarity between the input image and pre-trained feature maps

Show answer

Correct answer

B
By computing the gradient of the model’s output with respect to the input image

Question 14 on its own page

Question 15

+2 marksOne correct option

Consider an object detection system evaluated on a dataset consisting of 1000 images. The system makes 1500 predictions across these images, and for each image, there are annotated ground truth bounding boxes.
The system’s precision-recall curve is calculated, and the precision at different recall levels for one of the classes is as follows:

Recall	Precision
0.1	0.90
0.2	0.85
0.3	0.80
0.4	0.75
0.5	0.70
0.6	0.65
0.7	0.60
0.8	0.55
0.9	0.50
1.0	0.45

Calculate the mean Average Precision (mAP) for this class using the 11-point interpolation method, which averages the precision values at recall levels {0.0, 0.1, 0.2, ..., 1.0}. The precision at recall 0.0 can be assumed to be 1.0.
Additionally, the system’s AP values for the other two classes are as follows:
- AP for class 2: 0.78
- AP for class 3: 0.72
Based on these AP values, what is the mean Average Precision (mAP) across all three classes?

A
0.70
B
0.69
C
0.73
D
0.76

Show answer

Correct answer

C
0.73

Question 15 on its own page

Question 16

+2 marksOne correct option

Which one of the following statements is false? (Pick the most appropriate one.)

A
Attention mechanisms cannot be applied to the bidirectional RNN model
B
An image captioning network cannot be trained end-to-end even though we are using 2 different modalities to train the network
C
One of the key components in the vanilla transformer is the recurrent connection that help them to deal with variable input length.
D
All of these

Show answer

Correct answer

D
All of these

Question 16 on its own pageAlso asked in End Term 22 Dec 2024, Quiz 2 3 Aug 2025, Quiz 2 23 Nov 2025

Question 17

+2 marksOne or more correct options

Which of the following statements is True (select all that apply):

Select all that apply.

A
The number of learnable parameters in an RNN grows exponentially with input sequence length considered
B
Long sentences give rise to the vanishing gradient problem
C
Electrocardiogram signal classification is a sequence learning problem
D
RNNs can have more than one hidden layer

Show answer

Correct answers

B
Long sentences give rise to the vanishing gradient problem
C
Electrocardiogram signal classification is a sequence learning problem
D
RNNs can have more than one hidden layer

Question 17 on its own page

Question 18

+2 marksOne or more correct options

Which of the following is true regarding Hard Attention and Soft Attention?

Select all that apply.

A
Soft Attention is smooth and differentiable
B
Variance reduction techniques are used to train Soft Attention models
C
Soft Attention is computationally cheaper than Hard Attention when the source input is large
D
The inference (test time) overhead is low in Hard Attention when compared to Soft Attention models

Show answer

Correct answers

A
Soft Attention is smooth and differentiable
D
The inference (test time) overhead is low in Hard Attention when compared to Soft Attention models

Question 18 on its own pageAlso asked in Quiz 2 1 Dec 2024, Quiz 2 23 Nov 2025

Question 19

+2 marksOne or more correct options

Which of the following are true?

Select all that apply.

Show answer

Correct answers

Question 19 on its own page

Question 20

+3 marksNumerical answer

Consider a Faster R-CNN object detection framework, which is being trained with a minibatch size of 2. The RPN produces 512 proposals, out of which 256 and 128 of them are foreground proposals, corresponding to each image in the mini-batch. For a single mini-batch update, the bounding box regression loss will be computed _______________ times.

Show answer

Correct answer: 384

Question 20 on its own page

Question 21

+1 markNumerical answer

Based on the above data, answer the given subquestions.

Show answer

Correct answer: 4

Question 21 on its own page

Question 22

+1 markNumerical answer

Based on the above data, answer the given subquestions.

Show answer

Correct answer: 6

Question 22 on its own page

Question 23

+1 markNumerical answer

Based on the above data, answer the given subquestions.

Show answer

Correct answer: 3

Question 23 on its own page

Question 24

+1 markNumerical answer

Based on the above data, answer the given subquestions.

Show answer

Correct answer: 1

Question 24 on its own page

Question 25

+1 markNumerical answer

Based on the above data, answer the given subquestions.

Show answer

Correct answer: 2

Question 25 on its own page

Question 26

+1 markNumerical answer

Based on the above data, answer the given subquestions.

Show answer

Correct answer: 5

Question 26 on its own page

Question 27

+3 marksNumerical answer

Based on the above data, answer the given subquestions.

Number of Parameters: __________________

Show answer

Correct answer: 12800

Question 27 on its own page

Question 28

+3 marksNumerical answer

Based on the above data, answer the given subquestions.

Computational Cost: __________________

Show answer

Correct answer: 209715200

Question 28 on its own page

Question 29

+1.5 marksNumerical answer

Let input have size $D_f \times D_f \times M$ where $D_f = 128$ and $M = 16$ and output feature map (after passing input through conv layer) has $D_f \times D_f \times N$ size where $N = 32$ . Assume padded convolution. Let width of the square kernel in conv layer be $k$ where $k = 5$ (Ignore the bias term in the calculation).

Calculate the number of parameters and computational cost, but make use of Depthwise Seperable convolution and pointwise convolution instead of standard convolution.

Based on the above data, answer the given subquestions.

Number of parameters for depthwise convolution: __________________

Show answer

Correct answer: 400

Question 29 on its own page

Question 30

+1.5 marksNumerical answer

Calculate the number of parameters and computational cost, but make use of Depthwise Seperable convolution and pointwise convolution instead of standard convolution.

Based on the above data, answer the given subquestions.

Computational Cost for depthwise convolution: ________________

Show answer

Correct answer: 6553600

Question 30 on its own page

Question 31

+1.5 marksNumerical answer

Calculate the number of parameters and computational cost, but make use of Depthwise Seperable convolution and pointwise convolution instead of standard convolution.

Based on the above data, answer the given subquestions.

Number of parameters for pointwise convolution: _________________

Show answer

Correct answer: 512

Question 31 on its own page

Question 32

+1.5 marksNumerical answer

Calculate the number of parameters and computational cost, but make use of Depthwise Seperable convolution and pointwise convolution instead of standard convolution.

Based on the above data, answer the given subquestions.

Computational cost for pointwise convolution: ___________________

Show answer

Correct answer: 8388608

Question 32 on its own page

Question 33

+1.2 marksNumerical answer

Based on the above data, answer the given subquestions.

Number of weights in Weight Matrix U1 is: ______________

Show answer

Correct answer: 5000

Question 33 on its own page

Question 34

+1.2 marksNumerical answer

Based on the above data, answer the given subquestions.

Number of weights in Weight Matrix V1 is: _______________

Show answer

Correct answer: 2500

Question 34 on its own page

Question 35

+1.2 marksNumerical answer

Based on the above data, answer the given subquestions.

Number of weights in Weight Matrix U2 is: ________________

Show answer

Correct answer: 10000

Question 35 on its own page

Question 36

+1.2 marksNumerical answer

Based on the above data, answer the given subquestions.

Number of weights in Weight Matrix V2 is: ______________

Show answer

Correct answer: 40000

Question 36 on its own page

Question 37

+1.2 marksNumerical answer

Based on the above data, answer the given subquestions.

Number of weights in Weight Matrix W is: ________________

Show answer

Correct answer: 200000

Question 37 on its own page