Computer Vision Exam Questions

Based on Lectures 6, 7, 8, 9

I. Essay Questions

Lecture 6: Image Recognition

1. Explain in detail the architecture of Convolutional Neural Networks (CNN) and

the role of each layer in the image classification process. (4 marks)

2. Compare between different pooling layers (Max pooling, Average pooling,

Global average pooling) in terms of mechanism, usage, and impact on network
performance. (3 marks)

3. Explain the problems of Overfitting and Underfitting in the context of neural

networks, and how Dropout layers can be used to overcome one of them. (4
marks)

4. Describe the steps of training using Backpropagation in convolutional neural

networks, explaining how weights are updated during the process. (4 marks)

Lecture 7: Advanced CNN Architectures

1. Compare between AlexNet and VGGNet in terms of architecture, innovations,

and performance, explaining why VGGNet is considered an important
development in computer vision. (4 marks)

2. Explain the concept of the Inception Module in GoogLeNet, and how it

contributed to improving the performance of convolutional neural networks.
(3 marks)

3. Describe the concept of Residual Blocks in ResNet, and how Skip connections
help solve the Vanishing Gradient problem. (4 marks)

4. Explain the concept of Transfer Learning and its importance in computer

vision, clarifying the difference between using a pretrained network as a
classifier and using it as a feature extractor. (4 marks)
Lecture 8: Object Detection

1. Explain the difference between Image Classification and Object Detection,

highlighting the additional challenges in object detection. (3 marks)

2. Describe the concept of Region Proposals in object detection, and how this
technique evolved from Selective Search to Region Proposal Networks (RPN).
(4 marks)

3. Explain the mechanism of the Non-Maximum Suppression algorithm and its

role in improving object detection accuracy. (3 marks)

4. Compare between different object detector evaluation metrics (IOU, MAP,

Precision-Recall Curve), explaining how they are used to evaluate the
performance of object detection models. (4 marks)

Lecture 9: Object Detection (continued)

1. Explain the evolution from R-CNN to Fast R-CNN to Faster R-CNN, highlighting
the main improvements in each version and their impact on performance. (5
marks)

2. Compare between Two-stage Detection approach and One-stage Detection

approach, mentioning an example of each approach and their advantages and
disadvantages. (4 marks)

3. Explain the architecture of the SSD (Single Shot Detector) network and how it
differs from the Faster R-CNN architecture, highlighting the advantages and
disadvantages of each. (4 marks)

II. True or False Questions

Lecture 6: Image Recognition

1. Convolutional layers in CNNs are used for feature extraction while fully connected
layers are used for classification. (True)

2. Dropout layers are more effective when used in input layers compared to hidden
layers. (False)

3. Global Average Pooling calculates the average of all pixel values in the feature map,
which significantly reduces the network dimensions. (True)
 4. When using Backpropagation in training convolutional neural networks, the
weights of convolutional layers are updated before the weights of fully connected
layers. (False)

5. Underfitting is more difficult to address than Overfitting in deep neural networks.

(False)

Lecture 7: Advanced CNN Architectures

1. AlexNet uses ReLU activation function instead of the tanh function used in LeNet-5,
resulting in faster network training. (True)

2. VGGNet is characterized by using large filters (7×7 and 5×5) in the first
convolutional layers to extract more complex features. (False)

3. The Inception Module in GoogLeNet contains four parallel layers: a 1×1

convolutional layer, a 3×3 convolutional layer, a 5×5 convolutional layer, and a
3×3 pooling layer. (True)

4. Skip connections in ResNet improve information flow through the network, helping
to solve the Vanishing Gradient problem. (True)

5. When using Transfer Learning, all layers of the pretrained network must always be
retrained to achieve the best results. (False)

Lecture 8: Object Detection

1. Object Detection includes two main tasks: Object Localization and Object
Classification. (True)

2. Selective Search algorithm is faster and more accurate than Region Proposal
Network in generating regions of interest. (False)

3. Non-Maximum Suppression is used to remove overlap between bounding boxes

that detect the same object. (True)

4. Intersection Over Union (IOU) is more important than Mean Average Precision
(MAP) in evaluating the performance of object detection models. (False)

5. R-CNN networks use SVM classifier for classification while Fast R-CNN networks use
softmax layer. (True)
Lecture 9: Object Detection (continued)

1. Fast R-CNN applies the CNN feature extractor to each region of interest (RoI)
separately, which improves detection speed compared to R-CNN. (False)

2. Faster R-CNN uses Region Proposal Network (RPN) as an attention network to

guide the network's attention to interesting regions in the image. (True)

3. One-stage Detectors like SSD are slower but more accurate than Two-stage
Detectors like Faster R-CNN. (False)

4. SSD uses multi-scale feature layers to detect objects of different sizes. (True)

5. All modern object detection models are trained from scratch without relying on
pretrained networks for feature extraction. (False)

III. Multiple Choice Questions

Lecture 6: Image Recognition

1. Which of the following hyperparameters in convolutional layers determines

the number of convolutional filters in each layer?
2. a) Kernel size
3. b) Stride
4. c) Padding

5. d) Depth

6. What is the main purpose of using pooling layers in convolutional neural

networks?

7. a) Increasing network depth

8. b) Reducing network size and number of parameters
9. c) Increasing classification accuracy

10. d) Improving forward propagation speed only

11. What problem does the Dropout Layer address in neural networks?

12. a) Underfitting
13. b) Vanishing Gradient
14. c) Overfitting

15. d) Exploding Gradient

 16. What is the first step in training convolutional neural networks using
Backpropagation?

17. a) Calculating the total error at the output layer

18. b) Initializing all filters and parameters/weights with random variables
19. c) Updating all filter values/weights
20. d) Calculating the gradients of the error with respect to all weights in the network

Lecture 7: Advanced CNN Architectures

1. What was the main innovation introduced by AlexNet compared to LeNet-5?

2. a) Using pooling layers only
3. b) Using tanh activation function
4. c) Using ReLU activation function and Dropout layer

5. d) Using 3×3 filters only

6. What is the main advantage of using 3×3 filters in VGGNet instead of larger
filters?

7. a) Increasing the number of training parameters

8. b) Reducing the number of training parameters while maintaining the same
receptive field
9. c) Reducing network depth

10. d) Increasing training speed only

11. Which main component in ResNet helps solve the Vanishing Gradient problem?

12. a) Deep pooling layers

13. b) Inception Module
14. c) Skip connections

15. d) Multiple Dropout layers

16. Which of the following Transfer Learning approaches involves freezing some
layers of the pretrained network and training both the non-frozen layers and
the newly added classifier layers?

17. a) Using pretrained network as a classifier

18. b) Using pretrained network as a feature extractor
19. c) Fine-tuning
20. d) Training from scratch
Lecture 8: Object Detection

1. What is the main difference between Image Classification and Object

Detection?
2. a) Image Classification requires more data for training
3. b) Object Detection involves localizing objects in addition to classifying them
4. c) Image Classification uses neural networks only while Object Detection uses
traditional algorithms

5. d) Object Detection is computationally simpler than Image Classification

6. What is the purpose of the Non-Maximum Suppression algorithm in object

detection?

◦ a) Increasing the number of bounding boxes for each object

◦ b) Reducing processing time by ignoring some regions
◦ c) Ensuring the algorithm detects each object only once
◦ d) Improving classification accuracy only

7. What is Intersection Over Union (IOU)?

◦ a) The ratio of correct predictions to total predictions

◦ b) A measure of overlap between the predicted bounding box and the ground
truth bounding box
◦ c) The number of frames that can be processed per second
◦ d) The ratio of detected objects to the total number of objects in the image

8. What is the first step in the R-CNN algorithm for object detection?

◦ a) Applying CNN to the entire image

◦ b) Using Selective Search algorithm to propose about 2,000 regions
◦ c) Classifying each region using softmax layer
◦ d) Applying Non-Maximum Suppression (NMS)

Lecture 9: Object Detection (continued)

1. What was the main improvement introduced by Fast R-CNN compared to the
original R-CNN?

◦ a) Using Region Proposal Network (RPN) instead of Selective Search

algorithm
◦ b) Applying CNN to the entire image first and then proposing regions
◦ c) Using RoI pooling layer to extract a fixed-size window
◦ d) Using softmax layer instead of SVM classifier
 2. What is the main difference between Fast R-CNN and Faster R-CNN?

◦ a) Faster R-CNN uses Region Proposal Network (RPN) instead of Selective

Search algorithm
◦ b) Fast R-CNN is faster but less accurate
◦ c) Faster R-CNN uses SVM classifier instead of softmax layer
◦ d) Fast R-CNN uses multi-scale feature layers

3. What is the main advantage of One-stage Detectors like SSD compared to Two-
stage Detectors like Faster R-CNN?

◦ a) Higher accuracy in all cases

◦ b) Higher speed at the expense of some accuracy
◦ c) Lower memory consumption only
◦ d) Better ability to detect small objects

IV. Fill in the Blank Questions

Lecture 6: Image Recognition

1. The general architecture of Convolutional Neural Networks (CNN) consists of an _

layer, _ layers for feature extraction, _ layers for classification, and an _ layer for
prediction.

2. The four hyperparameters in convolutional layers are _ which represents the

number of convolutional filters, _ which represents the size of the convolutional
filter matrix, _ which represents the amount by which the filter slides over the
image, and _ which adds zeros around the border of an image.

3. Pooling layers help reduce the _ of the network by reducing the number of _
passed to the next layer, and there are two main types of pooling layers: _ and _.

4. _ occurs when the model fails to fit the training data, while _ occurs when the
model memorizes the training data instead of learning the actual features.

Lecture 7: Advanced CNN Architectures

1. _ network was developed by Lecun and his team in 1998, and it uses _
activation function instead of the ReLU function used currently.

2. _ network won the ILSVRC image classification competition in 2012, and

introduced innovations such as using _ activation function and ____ layer with a
probability of 0.5 in the fully connected layers.
 3. _ network, also known as VGG16, is characterized by using _ filters instead of
larger filters, and consists of 16 weight layers: _ convolutional layers and _ fully
connected layers.

4. The Inception Module in _ network consists of four layers: a _ convolutional layer,

a _ convolutional layer, a _ convolutional layer, and a ____ pooling layer.

5. Transfer Learning means transferring what a neural network has learned from
being trained on a specific _ to another related problem, and there are three
major transfer learning approaches: _, _, and _.

Lecture 8: Object Detection

1. Object Detection is a computer vision task that involves two main tasks: _ and _ of
objects within an image.

2. Object detection methods are divided into two types: object detection using _ and
object detection using _.

3. The general object detection framework consists of four steps: _, _, _, and _.

4. Non-Maximum Suppression (NMS) looks at all the boxes surrounding an object to

find the box that has the maximum ____, and it suppresses or eliminates the other
boxes.

5. Object detector evaluation metrics include _ to measure detection speed, _ to

measure network precision, _ to evaluate the overlap between bounding boxes,
and _ to evaluate model performance.

Lecture 9: Object Detection (continued)

1. The architecture of Fast R-CNN consists of the following modules: _ module to

extract features from the full image, _ to propose about 2,000 region candidates
per image, _ layer to extract a fixed-size window, and a _ output layer that
branches into two heads.

2. Faster R-CNN uses _ network instead of Selective Search algorithm, and this
network is known as an _ network because it guides the network's attention to
interesting regions in the image.

3. Detection approaches are divided into two types: _ detectors where detection
happens in two stages, and _ detectors that skip the region proposal stage and
run detection directly.
4. The architecture of the SSD model is composed of three main parts: _ to extract
feature maps, _ which is a series of convolution filters, and ____ used to eliminate
overlapping boxes and keep only one box for each object detected.