CVI FINAL EXAM DAY

Question 1

What is the key difference between RoI Pooling and RoI Align

Answer 1

RoI Pooling rounds coordinates causing misalignment while RoI Align uses interpolation to preserve spatial accuracy

Question 2

Why is SURF considered faster than SIFT

Answer 2

SURF uses approximations like box filters and integral images while SIFT uses precise gradients making it slower

Question 3

What kind of transformation does a Region Proposal Network RPN learn

Answer 3

It learns to transform anchor boxes into tighter object proposals by predicting shifts and scales

Question 4

Why is C3D not used for optical flow estimation

Answer 4

Because C3D performs video classification not pixelwise motion estimation so it does not output motion vectors

Question 5

How does DeepLab use atrous convolution and ASPP for segmentation

Answer 5

It uses atrous convolution to enlarge receptive field and ASPP to capture multiscale context features

Question 6

Fast RCNN

Answer 6

Deep learning for object detection

Question 7

Faster RCNN

Answer 7

Deep learning for object detection with learned region proposals

Question 8

RCNN

Answer 8

Deep learning for object detection

Question 9

OmniMotion

Answer 9

Advanced motion estimation using video tracking across occlusions

Question 10

Lucas Kanade

Answer 10

Traditional optical flow estimation

Question 11

Horn Schunck

Answer 11

Traditional optical flow estimation

Question 12

FlowNet

Answer 12

Deep learning for optical flow estimation

Question 13

VoxelMorph

Answer 13

Deep learning for image registration

Question 14

What does Otsus method actually optimise during threshold selection

Answer 14

It maximises inter class variance between foreground and background pixel distributions

Question 15

Name one deep learning model commonly used for image registration

Answer 15

VoxelMorph

Question 16

Prewitt

Answer 16

Traditional edge detection using gradient approximation

Question 17

AlexNet

Answer 17

Deep learning for image classification

Question 18

What is the difference between self supervised and unsupervised learning

Answer 18

Self supervised uses pseudo labels from data while unsupervised finds structure without labels

Question 19

Robinson

Answer 19

Traditional edge detection using compass kernels

Question 20

C3D

Answer 20

Deep learning for video classification and action recognition

Question 21

How do skip connections in U Net help during segmentation

Answer 21

They pass high resolution features from encoder to decoder to preserve localisation and detail

Question 22

HED

Answer 22

Deep learning for edge detection using holistically nested networks

Question 23

DoG

Answer 23

Traditional edge detection using Difference of Gaussians

Question 24

How does contrastive loss function mathematically encourage representation learning

Answer 24

It pulls similar pairs closer and pushes dissimilar pairs apart in embedding space using similarity scores

Question 25

How are eigenfaces constructed using PCA

Answer 25

By computing eigenvectors from face dataset and representing new faces as weighted sums of these eigenfaces

Question 26

When would you use mutual information instead of SSD for image registration

Answer 26

When the two images come from different modalities or have different intensity scales

Question 27

What is the visual output difference between semantic and panoptic segmentation

Answer 27

Semantic labels each pixel with a class while panoptic also distinguishes between different object instances

Question 28

UNet

Answer 28

Deep learning for semantic segmentation

Question 29

Mask RCNN

Answer 29

Deep learning for instance segmentation

Question 30

FlowNet2

Answer 30

Deep learning for optical flow estimation

Question 31

YOLO

Answer 31

Deep learning for real time object detection

Question 32

Sobel

Answer 32

Traditional edge detection using gradient approximation

Question 33

Kirsch

Answer 33

Traditional edge detection using compass kernels

Question 34

Canny

Answer 34

Traditional edge detection with gradient smoothing and hysteresis

Question 35

LoG

Answer 35

Traditional edge detection using Laplacian of Gaussian

Question 36

Rich Feature Hierarchies

Answer 36

Deep learning based edge detector built on top of CNNs

Question 37

TimeSformer

Answer 37

Transformer based model for video understanding

Question 38

DyeNet

Answer 38

Deep learning for video object segmentation using appearance and motion cues

Question 39

Two Stream CNN

Answer 39

Deep learning for action recognition using RGB and optical flow as parallel inputs

Question 40

How to calculate weights in CNN?

Answer 40

Kernel x kernel x colour channel number x filters

Question 41

How to calculate bias in CNN?

Answer 41

Number of filters

Question 42

What are two morphological operators and outline their purpose

Answer 42

Dilation – Expands the boundaries of foreground (white) regions. 🔹 Used to fill in small holes or connect nearby segmented regions. Erosion – Shrinks foreground regions by eroding boundaries. 🔹 Used to remove small noise or separate objects that are touching.

Question 43

What region proposal method does RCNN use

Answer 43

Selective Search

Question 44

How does RCNN extract features

Answer 44

It applies a CNN separately to each proposed region

Question 45

What is the output of RCNN

Answer 45

Class label and bounding box for each region

Question 46

What region proposal method does Fast RCNN use

Answer 46

Selective Search

Question 47

How does Fast RCNN extract features

Answer 47

It applies a CNN to the entire image once and uses RoI Pooling on the feature map

Question 48

What is the output of Fast RCNN

Answer 48

Class label and bounding box for each region

Question 49

What region proposal method does Faster RCNN use

Answer 49

Region Proposal Network RPN

Question 50

How does Faster RCNN extract features

Answer 50

It uses a CNN on the entire image and applies RoI Pooling to shared feature maps

Question 51

What is the output of Faster RCNN

Answer 51

Class label and bounding box for each region

Question 52

What region proposal method does Mask RCNN use

Answer 52

Region Proposal Network RPN

Question 53

How does Mask RCNN extract features

Answer 53

It uses a CNN on the entire image and applies RoI Align to shared feature maps

Question 54

What is the output of Mask RCNN

Answer 54

Class label bounding box and segmentation mask

Question 55

What is a feature map

Answer 55

The output of a convolutional layer in a CNN: It’s a 3D tensor: [Height × Width × Channels]. Each channel detects a specific feature (edge, curve, texture) Shows how much of X pattern is in each area

Question 56

What is spatial resolution

Answer 56

The size (in pixels) of the output feature map at any layer (decreases as you go deeper in a CNN) i.e. how big is the output grid

Question 57

What is a feature hierarchy

Answer 57

The idea that lower layers detect simple features (edges, corners), while higher layers detect complex features (faces, wheels, textures) Makes CNNs good at generalising across task.