Course outline
Module 1
Module 1: Foundations of Image Formation
Foundations of Image Formation
Knowledge check
Q1. In a digital grayscale image, what does each pixel typically store?
- A. A vector describing a geometric shape
- B. A single intensity value representing brightness ✓
- C. The RGB color triplet of a surface
- D. The depth of the scene from the camera
Why: A grayscale pixel stores one intensity value (commonly 0-255 for 8-bit images) that represents how bright that point is. Color images store multiple values per pixel, such as RGB triplets, but grayscale images use a single value.
Q2. What is the main role of the lens in a simple pinhole camera model?
- A. To amplify the light intensity at the sensor
- B. To store the captured image for later retrieval
- C. The pinhole itself replaces the lens, which is not needed in this model ✓
- D. To convert light into electrical signals
Why: The ideal pinhole camera has no lens: light from each scene point passes through a single tiny aperture and projects onto the image plane. Real cameras use lenses to gather more light while approximating this pinhole geometry.
Q3. Which quantity is most directly determined by the focal length of a camera?
- A. The color temperature of the image
- B. The field of view of the camera ✓
- C. The bit depth of each pixel
- D. The frame rate of video capture
Why: For a given sensor size, a shorter focal length produces a wider field of view, while a longer focal length narrows it. Focal length therefore controls how much of the scene is captured.
Q4. What does the term 'resolution' most commonly refer to in digital imaging?
- A. The number of pixels in the image grid ✓
- B. The total brightness of the image
- C. The number of colors available in the palette
- D. The speed at which the shutter opens
Why: Image resolution is typically expressed as the number of pixels along each dimension (for example, 1920x1080). More pixels allow finer spatial detail to be recorded, though resolution is distinct from brightness or color depth.
Q5. In a typical digital camera sensor, what physical process converts incoming light into an electrical signal?
- A. Chemical reactions in silver halide crystals
- B. The photoelectric effect in semiconductor photodiodes ✓
- C. Thermal expansion of a metal filament
- D. Magnetic induction in a coil
Why: CCD and CMOS sensors use photodiodes built from semiconductor material. When photons strike the material, they liberate electrons through the photoelectric effect, and the accumulated charge is read out as a signal proportional to light intensity.
Module 2
Module 2: Image Processing and Filtering
Image Processing and Filtering
Knowledge check
Q1. What is the primary purpose of a Gaussian blur filter in image processing?
- A. To detect edges in an image
- B. To reduce noise and smooth an image ✓
- C. To increase the contrast of an image
- D. To convert a color image to grayscale
Why: A Gaussian blur convolves the image with a Gaussian kernel, averaging nearby pixels with weights that fall off with distance. This smooths out high-frequency noise while preserving the overall structure of the image.
Q2. Which kernel size is most commonly used for a basic 3x3 mean (box) filter, and what does each output pixel become?
- A. The maximum value of the 3x3 neighborhood
- B. The median value of the 3x3 neighborhood
- C. The average of the 9 pixel values in the 3x3 neighborhood ✓
- D. The difference between the center pixel and its neighbors
Why: A mean filter replaces each pixel with the average of all pixels in its 3x3 neighborhood (9 values). This is a simple form of smoothing that reduces noise but can blur edges.
Q3. What is the main advantage of a median filter over a mean filter for removing salt-and-pepper noise?
- A. It is faster to compute
- B. It preserves edges better while removing extreme outlier pixel values ✓
- C. It always increases image brightness
- D. It works only on color images
Why: Salt-and-pepper noise creates extreme outlier pixels. The median is robust to such outliers, so it removes them without the averaging effect that would smear the noise into neighboring pixels and blur edges as a mean filter does.
Q4. In a Sobel edge detection operator, what does the gradient magnitude represent?
- A. The average brightness of the image
- B. The strength of intensity change at each pixel, indicating edge presence ✓
- C. The number of colors in the image
- D. The sharpness of the camera lens
Why: The Sobel operator approximates the image gradient in the x and y directions. The gradient magnitude, computed from these two components, is large where intensity changes rapidly, which corresponds to edges.
Q5. What happens to the spatial resolution of an image when you apply a convolution with a kernel and no padding?
- A. It always increases
- B. It stays exactly the same
- C. It decreases slightly, since border pixels cannot be fully centered on the kernel ✓
- D. It becomes three times larger
Why: Without padding, a kernel can only be centered on pixels at least half its size away from the border, so the output is smaller than the input. For example, a 3x3 kernel on a 100x100 image without padding yields a 98x98 output.
Module 3
Module 3: Feature Detection and Description
Feature Detection and Description
Knowledge check
Q1. What is the main purpose of feature detection in computer vision?
- A. To compress images for faster storage
- B. To identify distinctive points or regions in an image that can be reliably located across different views ✓
- C. To convert color images into grayscale
- D. To increase the resolution of an image
Why: Feature detection finds distinctive, repeatable points or regions (such as corners or blobs) that can be matched across images, which is essential for tasks like image stitching, tracking, and 3D reconstruction.
Q2. Which of the following is a classic corner detector introduced by Harris and Stephens?
- A. Canny edge detector
- B. Sobel operator
- C. Harris corner detector ✓
- D. Hough transform
Why: The Harris corner detector, proposed by Harris and Stephens in 1988, identifies corners by analyzing intensity changes in all directions using the structure tensor of image gradients.
Q3. What does a feature descriptor do?
- A. It detects the location of keypoints in an image
- B. It encodes the local appearance around a keypoint into a vector for matching ✓
- C. It removes noise from the image
- D. It classifies the object in the image
Why: A descriptor summarizes the local image neighborhood around a keypoint as a numerical vector, allowing keypoints to be compared and matched between images.
Q4. Which property makes SIFT features robust to changes in image scale?
- A. They use a fixed window size for all images
- B. They are detected across a scale-space pyramid and assigned a characteristic scale ✓
- C. They ignore all gradient information
- D. They are computed only on color channels
Why: SIFT builds a scale-space using Difference of Gaussians and selects keypoints at their characteristic scale, so the same feature can be recognized even when the image is zoomed in or out.
Q5. Why is ORB often preferred over SIFT in real-time applications?
- A. ORB produces more matches than SIFT in all cases
- B. ORB is computationally faster and uses binary descriptors compared by Hamming distance ✓
- C. ORB does not require keypoint detection
- D. ORB works only on 3D point clouds
Why: ORB is an efficient detector-descriptor combining FAST and BRIEF ideas, producing binary descriptors that can be matched very quickly with Hamming distance, making it suitable for real-time systems.
Module 4
Module 4: Camera Geometry and Calibration
Camera Geometry and Calibration
Knowledge check
Q1. What does the intrinsic matrix K in a pinhole camera model primarily describe?
- A. The position and orientation of the camera in the world
- B. The internal camera parameters such as focal length, principal point, and skew ✓
- C. The lens distortion coefficients only
- D. The transformation from world coordinates to image coordinates without any scaling
Why: The intrinsic matrix K contains internal parameters of the camera: focal lengths (fx, fy), the principal point (cx, cy), and skew. The extrinsic parameters, not K, describe camera position and orientation.
Q2. Which parameters are described by the extrinsic matrix [R|t] of a camera?
- A. Focal length and pixel size
- B. Principal point offset
- C. Rotation and translation relating world coordinates to camera coordinates ✓
- D. Radial and tangential distortion
Why: The extrinsic matrix [R|t] consists of a 3x3 rotation matrix R and a translation vector t that transform points from the world coordinate frame into the camera coordinate frame.
Q3. What is the main purpose of camera calibration in computer vision?
- A. To increase the resolution of captured images
- B. To determine the camera's intrinsic and extrinsic parameters for accurate geometric measurements ✓
- C. To compress images for faster transmission
- D. To convert color images to grayscale
Why: Camera calibration estimates intrinsic and extrinsic parameters so that pixel coordinates can be accurately related to 3D world coordinates, enabling measurement, 3D reconstruction, and pose estimation.
Q4. A common calibration method uses a planar pattern such as a checkerboard. Why is a checkerboard popular for this task?
- A. Its colors make it easy to detect under any lighting
- B. Its corners are easy to detect precisely and their 3D positions are known ✓
- C. It automatically removes lens distortion
- D. It eliminates the need for multiple images
Why: Checkerboard corners are high-contrast, easy to localize with sub-pixel accuracy, and their real-world spacing is known, providing reliable 2D-3D correspondences for calibration.
Q5. What does radial distortion in a camera lens cause?
- A. Straight lines in the scene appearing curved in the image ✓
- B. Colors appearing shifted between channels
- C. The image becoming darker toward the edges only
- D. Objects appearing blurred due to motion
Why: Radial distortion, such as barrel or pincushion distortion, bends straight lines in the world into curves in the image, most noticeably away from the image center. Calibration estimates coefficients to correct it.
Module 5
Module 5: Stereo Vision and Depth Estimation
Stereo Vision and Depth Estimation
Knowledge check
Q1. In stereo vision, what is the term for the horizontal shift of a point's position between the left and right images?
- A. Baseline
- B. Disparity ✓
- C. Focal length
- D. Epipole
Why: Disparity is the difference in the horizontal image coordinates of the same 3D point as seen in the left and right images. Larger disparity indicates the point is closer to the cameras.
Q2. What is the relationship between depth Z and disparity d for a rectified stereo pair with focal length f and baseline B?
- A. Z = f * d / B
- B. Z = f * B / d ✓
- C. Z = d * B / f
- D. Z = f + B + d
Why: Depth is inversely proportional to disparity: Z = f * B / d. As disparity increases, depth decreases, which is why nearby objects appear to shift more between the two views.
Q3. What does the baseline in a stereo camera setup refer to?
- A. The distance from the cameras to the scene
- B. The distance between the optical centers of the two cameras ✓
- C. The resolution of the camera sensors
- D. The angle between the image planes
Why: The baseline is the distance between the optical centers of the two cameras. A larger baseline generally gives better depth precision for distant objects, since it produces larger disparities.
Q4. What is the main purpose of rectifying a stereo image pair?
- A. To increase the resolution of the images
- B. To remove color differences between cameras
- C. To align the epipolar lines horizontally so correspondence search is one-dimensional ✓
- D. To convert the images to grayscale
Why: Rectification transforms the image pair so that corresponding points lie on the same horizontal scanline. This reduces the stereo correspondence search from a 2D problem to a 1D search along each row.
Q5. Which of the following is a common challenge in stereo matching?
- A. Textureless regions, where matching points is ambiguous ✓
- B. Too many color channels in the images
- C. Images being captured at too high a frame rate
- D. Cameras having identical focal lengths
Why: Textureless regions such as plain walls or sky provide few distinctive features, making it hard to find reliable correspondences between the left and right images. Occlusions and repetitive patterns are also common challenges.
Module 6
Module 6: Object Detection and Recognition
Object Detection and Recognition
Knowledge check
Q1. What is the primary goal of object detection in computer vision?
- A. To classify an entire image into a single category
- B. To locate and classify multiple objects within an image using bounding boxes
- C. To enhance the resolution of an image
- D. To convert a color image into grayscale
Q2. What does a bounding box typically consist of in object detection?
- A. A set of pixel intensity values
- B. A mask outlining the exact shape of an object
- C. Coordinates describing a rectangle that encloses an object
- D. A histogram of color distributions
Q3. What is the difference between object detection and image classification?
- A. Classification identifies object locations; detection does not
- B. Detection assigns a single label to the whole image; classification locates objects
- C. Classification assigns a label to an image; detection both locates and labels objects
- D. There is no difference; the terms are interchangeable
Q4. What metric is commonly used to measure the overlap between a predicted bounding box and a ground-truth box?
- A. Mean Squared Error
- B. Intersection over Union (IoU)
- C. Cross-entropy loss
- D. Peak Signal-to-Noise Ratio
Q5. What is Non-Maximum Suppression (NMS) used for in object detection?
- A. To remove noise from the input image
- B. To increase the brightness of detected objects
- C. To eliminate redundant overlapping bounding boxes for the same object
- D. To resize images before feeding them to a network
Module 7
Module 7: Deep Learning for Vision
Deep Learning for Vision
Knowledge check
Q1. What is the main purpose of a convolutional layer in a convolutional neural network (CNN)?
- A. To randomly drop neurons during training to prevent overfitting
- B. To slide learnable filters across the input to detect local spatial features such as edges and textures ✓
- C. To convert the image into a one-dimensional vector for classification
- D. To normalize pixel values to a fixed range between 0 and 1
Why: Convolutional layers apply small learnable filters (kernels) across the input, producing feature maps that capture local patterns like edges, corners, and textures. Sharing these filters across the image also makes the network efficient and translation-equivariant.
Q2. What does a pooling layer, such as max pooling, primarily do in a CNN?
- A. It increases the spatial resolution of the feature maps
- B. It adds non-linearity to the network's outputs
- C. It reduces the spatial dimensions of feature maps while retaining the most salient information ✓
- D. It learns new filters from the training data
Why: Pooling downsamples feature maps, for example by taking the maximum value in each local window. This reduces computation and the number of parameters and gives the network some tolerance to small shifts in the input.
Q3. Which activation function is most commonly used in hidden layers of modern CNNs?
- A. Sigmoid
- B. Softmax
- C. ReLU (Rectified Linear Unit) ✓
- D. Linear identity function
Why: ReLU, defined as max(0, x), is widely used because it is simple to compute, helps mitigate vanishing gradients compared to sigmoid, and induces sparse activations. Softmax is typically reserved for the output layer in multi-class classification.
Q4. What is the purpose of data augmentation when training an image classification model?
- A. To increase the image resolution so that the model sees more detail
- B. To artificially expand the training set with transformed versions of images, improving generalization ✓
- C. To remove noisy or mislabeled samples from the dataset
- D. To speed up inference time on new images
Why: Data augmentation applies transformations such as flips, rotations, crops, and color jitter to create varied training examples. This helps the model become robust to variations it will encounter in real data and reduces overfitting.
Q5. What is transfer learning in the context of deep learning for vision?
- A. Moving a trained model from one GPU to another
- B. Reusing a model pretrained on a large dataset, such as ImageNet, as a starting point for a new but related task ✓
- C. Converting images from RGB to grayscale before training
- D. Training a model using data from multiple unrelated domains at once
Why: Transfer learning leverages features learned on a large source dataset, often ImageNet, and adapts them to a new task with less labeled data and training time. Early layers capture general features such as edges, which transfer well across tasks.