Computer Vision: Midterm Study Guide

Instructor: V. Paúl Pauca

1 Concepts
Here are the concepts I expect you to be familiar with and what you should be able to do with them. Review
these concepts using the course slides, your notes, the associated parts of textbook, and your project work.

1. Digital images as 2D signals, sampling and quantization

• treat images as 2D functions in space and describe how intensity is sample and quantized
• describe how color is represented (RGB, YUV, HSV)

2. Image filtering, point-wise operators, linear and non-linear filters, image boundaries
You should be able to

• describe at least one point-wise operator, such as contrast stretching

3. Nonlinear filters and image boundaries

• describe how the median filter is applied and why it is said to be non-linear
• describe how the filtering process deals with image boundaries and what methods are commonly
used

4. The convolution operation, the discrete Fourier transform and the convolution theorem
You should be able to

• write down the convolution operation (the same as the equation for linear filtering)
• write down the forward and inverse 2D discrete Fourier transforms and identify the terms and
point out their outputs (i.e. what does the forward 2D DFT provides? what does te inverse 2D
DFT provides?)
• specify the convolution theorem and outline the steps in which the theorem is used to transform
a 2D image signal.

5. Analyzing images by their frequency content

• describe frequency content of an image from the plot of its power spectra in a log +1 plot.
• describe what the low, mid, and high frequencies of an image are and the type of information
they encode
CSC 391
Spring 2019
1
 • identify low and high frequencies in a 2D fftshifted plot of the power spectra
• infer the frequency content produced by edges
6. Filtering in the frequency domain
You should be able to
• explain the difference between box and Gaussian filtering using the convolution theorem
• describe why an ideal filter produces artifacts in an image and propose methods to solve this issue
(i.e., Butterworth filtering)
• define highpass filters from lowpass filters
7. Sampling and aliasing
You should be able to
• describe what aliasing is and how it may occur during the sampling and subsampling process of
a 2D signal
• give examples of aliasing in 2D images
• describe how aliasing can be avoided or minimized during subsampling of a 2D signal
8. Image filtering as correlation, image pyramids
You should be able to:
• describe how to compute the correlation response of an image with to a given path, using zero-
mean correlation, sum of squared differences, and normalized cross-correlation
• evaluate the correlation response and its limitations, i.e., scale difference between image and patch
• Understand the concept of image pyramids and know how to build Gaussian and Laplacian pyra-
mids
9. Derivative filters, image gradients, edge detection under noise, the Canny edge detection algorithm
You should be able to
• define what an edge is and specify the conditions causing edges in a digital image
• show the result of applying a derivative filter to a 1D signal that may or may not contain noise
• show how smoothing prior to edge detection deals with noisy signals
• draw the convolution of a 1D signal with a filter (e.g., Gaussian + derivative)
• specify the different steps of the Canny edge detection algorithm
10. Feature extraction, corner detection
You should be able to:
• Describe and discuss the characteristics of good image features, such as repeatability, saliency,
compactness and efficiency, and locality
• Identify and contrast the different characteristics afforded by edges vs corners
• List at least three applications of image features
• Describe the idea behind corner detection via the sum of square differences within image windows
• Writer the second moment matrix and describe the information it encodes
• Describe the information conveyed by the eigenvalues of the second moment matrix
• Describe how the eigenvalues of the second moment matrix are used to define a corner response
function
• The performance of the Harris corner detection relative to intensity, translation, rotation, and
scale changes
CSC 391
Spring 2019
2
11. Feature extraction, blob detection
You should be able to
• Identify and contrast the characteristics of blob features versus corners
• Describe the basic idea behind blob detection, i.e., correlation with a “blob” filter over multiple
scales
• Specify how a blob filter is created mathematically using second partial derivatives of a Gaussian
• Show in 1D how the Laplacian of a Gaussian performs around edges
• Show in 1D how a blob (defined as the superposition of two edges) responds to a LoG of a given
scale (σ) and illustrate how the LoG response is maximum when the scale of the LoG and the
image feature match
• Describe why scale normalization by σ 2 is necessary in blob detection
• Determine the scale of a feature from its maximum blob response
12. Feature descriptors
You should be able to
• describe how transformations between two image regions are made to match their scale
• understand a descriptor obtained from a histogram of local gradient directions in an image patch
• understand the descriptors provided by the SIFT algorithm
• Evaluate the performance of SIFT descriptors relative to changes in intensity, rotation, translation,
and scale

CSC 391
Spring 2019
3