Edge detection Flashcards
What is an edge in Computer Vision?
A place in the image where the pixel value changes sharply and has discontinuities
Name two derivative filters
Prewitt ([1, 0, -1], [1, 0, -1], [1, 0, -1]) and
Sobel ([1, 0, -1], [2, 0, -2], [1, 0, -1]).
How do you calculate the gradient direction?
arctan(g_y, g_x)
Explain the Sobel filter as 2 1D filters
[1,2,1], smoothing and [1, 0, -1], derivative.
Why do we use the derivate of Gaussians
To remove noise before derivation and detect structures at different scales.
How does sigma affect the results when we take the derivative of Gaussians?
Larger sigma removes more noise and detects larger structures.
What is the formula for the Laplacian?
Lap = d^2f/d^2x + d^2f/d^2y
Why do we use the Guassian before laplacian (LoG)
The Laplacian is even more sensitive to noise than the derivatives.
What is the “Mexican hat” function
A flipped LOG (negative LoG)
Name the three conditions Canny proposed for a good edge detector
1) Good detection; should detect all edges
2) Good localization; should detect edges where they are
3) A single response, should only detect edges where they are
Describe Canny’s algorithm for edge detection
1) Gaussian filtering
2) Calculate gradient magnitude and direction
3) Perform non-maximum repression
4) Perform hysteresis thresholding
What its NMR, non-maximum repression in Canny’s?
For each pixel, we set the pixel to zero if it isn’t the maxima along the gradient direction in a neighborhood.
What its hysteresis thresholding in Canny’s?
Use two thresholds. The pixels above the upper threshold are always considered edges, pixels below the lower threshold are never considered edges.
Pixels between the thresholds are considered edges if they are connected to an edge.
How does the choice of sigma affect the found edges in Canny’s?
Small sima will detect small structures, large sigma will detect larger structures.
How can NMS be used in object detection?
If object regions have significant overlap only the strongest response is kept.
What is data-driven edge detection
Training machine learning algorithms to detect edges
What is the main disadvantage of data-driven edge detection?
It requires ground truth for all training data.
Name one advantage of data-driven edge detection over Canny´s.
It can detect structures over several scales at once, while the scale for canny is determined by sigma.
What is the double intercept form for a line?
x/a + y/b = 0, a is x-intercept, b is y-intercept.
What is the normal form for a line
xcos(theta) + ysin(theta) = p
What is the idea of hough-transform
We transform from the image space to parameter space (n,m or a,b or theta,p). The intercept points of lines in this space describe lines in the normal space.
What is the main advantage of the normal form over the slope and double intercept forms?
theta is limited to [0, pi] while the other parameters might be in the range [-inf, inf].
