CVI - ExamQuestions I should know

Question 1

What is the first step of the Canny edge detection algorithm?

Answer 1

Perform Gaussian filtering to suppress noise.

Question 2

How does the choice of sigma (𝜎) in Gaussian filtering affect Canny edge detection results?

Answer 2

A small 𝜎 detects fine features; a large 𝜎 detects large scale edges.

Question 3

What is the second step of the Canny edge detection algorithm?

Answer 3

Calculate the gradient magnitude and direction.

Question 4

How is the gradient magnitude (M) at a pixel calculated in Canny edge detection?

Answer 4

Using the derivatives along X (Gx) and Y (Gy), the magnitude is calculated as M = sqrt(Gx^2 + Gy^2).

Question 5

What is the third step of the Canny edge detection algorithm?

Answer 5

Apply non-maximum suppression (NMS) to get a single response for each edge.

Question 6

How does Non-Maximum Suppression (NMS) in Canny edge detection check if an edge is valid?

Answer 6

NMS checks if the gradient magnitude at a pixel is a local maximum along the gradient direction, aiming to get a single pixel‑wide response for each edge.

Question 7

What is the fourth step of the Canny edge detection algorithm?

Answer 7

Perform hysteresis thresholding to find potential edges.

Question 8

Explain hysteresis thresholding in Canny edge detection.

Answer 8

It uses two thresholds, t_high and t_low. Pixels with magnitude > t_high are accepted as edges. Pixels < t_low are rejected. Pixels between t_low and t_high are accepted only if they are connected to an accepted edge (>t_high).

Question 9

What is a common relationship between the high and low thresholds in Canny hysteresis thresholding?

Answer 9

Typically, t_high = 2 * t_low.

Question 10

According to the sources how are Canny and Sobel described in the context of edge detection?

Answer 10

Canny is described as a standard edge detector, while Sobel is listed as a typical operator for edge detection.

Question 11

What is an advantage of the Canny edge detector compared to the Sobel operator?

Answer 11

Canny generally produces thinner and more accurate edges and is less sensitive to noise due to its multi‑stage process.

Question 12

What is a disadvantage of the Canny edge detector compared to the Sobel operator?

Answer 12

Canny is computationally more complex and requires tuning multiple parameters (Gaussian sigma, high/low thresholds).

Question 13

What is an advantage of the Sobel operator compared to the Canny edge detector?

Answer 13

Sobel is simpler and faster to compute.

Question 14

What is a disadvantage of the Sobel operator compared to the Canny edge detector?

Answer 14

Sobel is more susceptible to noise and produces thicker edge responses.

Question 15

How can the Sobel operator contribute to detecting diagonal edges?

Answer 15

The Sobel operator calculates intensity gradients in the X and Y directions. The resulting gradient magnitude (sqrt(Gx^2 + Gy^2)) is high at pixels where intensity changes rapidly, regardless of orientation, thus highlighting diagonal edges.

Question 16

Do the sources mention other edge detection operators besides Sobel and Canny that can detect edges at different orientations?

Answer 16

Yes, typical operators include Prewitt, Sobel, Robinson, and Kirsch operators.

Question 17

Can a Convolutional Neural Network (CNN) be used for image classification?

Answer 17

Yes, Deep Learning models, including CNNs, are widely applied to tasks like image classification.

Question 18

What do CNNs need to perform final image classification?

Answer 18

A typical classification network requires fully connected layers and a classification layer (like Softmax) at the end.

Question 19

What is the formula provided for calculating the output dimension of a convolutional layer?

Answer 19

The formula is Sout = Floor((Sin + 2*padding - dilation*(kernel_size - 1) - 1) / stride) + 1.

Question 20

How can the network Conv1 -> Conv2 -> Pool1 -> Conv3 be modified to perform image classification with 10 classes?

Answer 20

The output of Conv3 needs to be flattened, then fed into one or more fully connected layers, followed by a classification layer (like Softmax) with 10 output units, one for each class.

Question 21

What is the primary source of information gathered for visual processing tasks?

Answer 21

Visual processing tasks use images or videos as input. Information gathering involves acquiring this visual data, typically via cameras.

Question 22

For a visual vehicle identification system aiming to identify vehicles by registration number what kind of information needs to be gathered and processed?

Answer 22

Images or video of vehicles must be captured. The raw data are pixel intensities, and key information to extract is the license plate text for identification.

Question 23

What techniques are typically used for edge detection?

Answer 23

Edge detection is often performed using image convolution with operators like Sobel, Prewitt, Robinson, or Kirsch, or using the multi‑stage Canny edge detection algorithm.

Question 24

What is a common method for detecting straight lines or circles from edge points?

Answer 24

The Hough Transform uses edge points to find parametric representations of shapes such as lines and circles.

Question 25

What are some traditional methods for **feature extraction** besides edge detection?

Answer 25

Beyond simple edge maps, classical feature extraction uses descriptors such as SIFT and SURF (built around corners/blobs) and LBP (built around local texture patterns). These descriptors turn visual structures into numeric vectors that older ML algorithms can handle.

Question 26

How do Deep Learning image models perform **feature extraction**?

Answer 26

Deep learning models, such as **Convolutional Neural Networks (CNNs)**, learn a hierarchy of features directly from the input data.

Question 27

What are some Deep Learning architectures mentioned for **Image Classification**?

Answer 27

Architectures include **LeNet, AlexNet, Inception (GoogLeNet), VGGNet, and ResNet**.

Question 28

What are some CNN-based methods for **Object Detection**?

Answer 28

Methods include **R‑CNN, Fast R‑CNN, Faster R‑CNN, and YOLO (You Only Look Once)**.

Question 29

What is **Semantic Segmentation** and what are examples of DL architectures for it?

Answer 29

Semantic Segmentation assigns a class label to each pixel ('stuff'). DL architectures include **U‑Net** and the **DeepLab** family (v2, v3, v3+).

Question 30

What is **Instance Segmentation** and what is an example method?

Answer 30

Instance Segmentation detects and masks each individual object instance ('things'). **Mask R‑CNN** is a method for this task.

Question 31

What task combines Semantic and Instance Segmentation?

Answer 31

**Panoptic Segmentation** combines pixel‑based segmentation ('stuff') with instance identification ('things').

Question 32

What is the goal of **Visual Saliency Modelling**?

Answer 32

It aims to identify regions of an image or video most likely to attract human attention, measuring how much something stands out.

Question 33

What are some applications of **Optical Flow** (calculating pixel motion)?

Answer 33

Applications include motion segmentation, video compression, structure‑from‑motion, and augmented reality.

Question 34

What Deep Learning models are mentioned for estimating **Optical Flow**?

Answer 34

Models like **FlowNet** and **FlowNet2** are used.

Question 35

What Deep Learning approaches are used for **Video Understanding** tasks like classification or action recognition?

Answer 35

Deep learning uses **Deep Features** and architectures like **TwoStream, C3D, and SlowFast networks**.

Question 36

What dimensionality reduction technique is mentioned?

Answer 36

**Principal Component Analysis (PCA)** reduces datasets by lowering dimensionality.

Question 37

What types of methods are used for **Image Registration**?

Answer 37

Methods include **intensity‑based methods** (MSD/SSD, Mutual Information) and **feature‑based methods** (using points like edges, corners).

Question 38

What are some types of **Generative Models**?

Answer 38

Types include **Generative Adversarial Networks (GANs)**, **Variational Autoencoders (VAEs)**, and diffusion models.

Question 39

What pre‑processing step is commonly used to **suppress noise** in images?

Answer 39

**Gaussian filtering** (smoothing) is commonly used to suppress noise.

Question 40

What are some low‑level vision tasks that can improve **low‑quality images**?

Answer 40

Tasks like **Denoising, Super‑resolution, and De‑blurring** can enhance image quality.

Question 41

What are some general challenges that make many computer‑vision tasks difficult?

Answer 41

Challenges include variations in **illumination, object pose, clutter, viewpoint, intra‑class appearance, and occlusions**.

Question 42

What is a drawback of edge detection and how is it mitigated?

Answer 42

Edge detection can be sensitive to **noise** (mitigated by smoothing such as Gaussian filtering) and can give fragmented edges (mitigated by NMS and hysteresis thresholding).

Question 43

What is a drawback of the Hough Transform for complex shapes?

Answer 43

Its **computational complexity can be high**, especially for shapes like circles that require a large parameter space.

Question 44

What is a major drawback for many supervised Deep Learning approaches?

Answer 44

They require **large datasets with expensive manual annotations**.

Question 45

How can the high data requirement for supervised Deep Learning be addressed?

Answer 45

Using **pre‑trained models, synthetic data, weak supervision, or semi‑supervised learning** can reduce the need for large annotated datasets.

Question 46

What issue arises in Semantic Segmentation due to downsampling layers?

Answer 46

Downsampling (e.g., pooling) can cause **reduced feature‑map resolution and localisation accuracy**.

Question 47

How are resolution and localisation issues addressed in Semantic Segmentation?

Answer 47

Techniques such as **atrous convolutions** and **skip connections** preserve or recover spatial detail.

Question 48

What problem occurs when applying models trained on one data type to another (e.g.images to video)?

Answer 48

A **domain shift** can lead to lower performance on the target domain.

Question 49

How can the problem of domain shift be addressed?

Answer 49

**Domain Adaptation** techniques improve performance when applying models across different data domains.

Question 50

What ethical concern is associated with Generative AI models?

Answer 50

They can be used to create **deepfakes and facilitate information manipulation**; responsible use is essential.