Detector Flashcards
What are detectors and why detectors might be better than segmentation
detectos allow decting on a specifc portions of the image meaning point while segmentation it just differentiates an object from that background
explain briefly corner detection
the idea consists of comparing the difference between two structuring element such that when we are on a continuous section we SSD Sum of the Square difference will be almost 0, on edge will be slightly higher than 0, while on corners very high from 0
How the image can be approximated by Taylor Series
For a given image I we have that
I(u+x,v+y) = the value at I(u,v) +Ix(u,v)x +Iy(u,v)y
where Ix(u,v)x represent the x derivative and Iy(u,v)y represents the y derivative consequnetly the SSD can be compute using the local deriative
Explain the Harris corner
The idea is that we would have for every pixel of the image a matrix representing the sum of the local derivative, hence, for eigen values L1, L2:
if L1=L2=0 = 0 -> continuous section
if L1=0 L2>0 -> border
if L1>0 L2>0 -> corner
since the local derivative is costly, a matrix representing the whole image, then we compute the local maxima
what are the pros and teh cons of Harris corner
for the pros : it can tolerate the rotations, the intensity of the grey level and image scale
However, it doesn’t tolerate the geometric invarience , hence the detection strongly relies on the size of the neighbourhood
what is taylor expansion
the idea is to approximate the sum of the squared difference using Taylor series which becomes the second derivative
how to avoid excessive computation of the matrix eigen
after the obtention of the matrix describing the for every pixel the local derivative, the eigenvalue is used to test the presence of corners, hence
this eigenvalues can summarize by the difference between the determinant of the array and K trace
when computing this alternative value for every pixel we can simply compute local maximum
How can we detect the scale in corner detection
the idea is to perform subsampling
explain the SIFT corner detection
for the scale :
for a given pyramid image (subsampling the image) , which can be performed by difference of gaussian , the scale is defined by sigma
the idea is to modify the resolution to extend the neighbourhood of where we look for a corner
then on every scale level we detect the harris corner accordingly, and we pick where have the best match
for the orientation :
we use the gradient and the magnitude at each point on the scale
explain how briefly how to use clustering and detectors to get some part of the cells work
the idea is that we first need to set up the region of interest, this can be done by detecting certain points, this can be done through harris corner, for instance, it depends on the problem we have
setup path descriptors we reflect dividing the image into several patches where each can describe for instance histogram .. etc
based on those obtained descriptors, we can compute the clustering for instance Kmean
how shi-Tomasi work
shi-Tomasi is similar to the harris corner however it relies on setting the minimum value between the two eigenvalues of the matrices obtained per pixel
what is the disadvantage of shi-Tomasi corner detector
it is not scaled invariant
explain the surf corner detector
Uses the integral image such we obtain for every point the sum up to that point,
this saves the excessive computation of the sum
once the integral image is obtained, we can perform convolutions at different scales efficiently
the advantage here is that we
explain the Fast corner detector
limited to 16 pixels around the center pixel, such that for a given region in image of a set consecutive pixels being dark then set of consecutive corner being bright are indications of a corner,h however it is not robust enough to scale
explain Fractal analysis with a Gaussian filter
the idea is to analyze the value of the image by a given guassian filter, such that some part of the images might be homogeneous which would be insensitive to the guassian filter which some other part will not be so
