max van kleek
6.869 advances in machine vision
problem set two
march 10, 2005

problem one

derivation a)

the difference-of-gaussian is a function DoG(x, y, sigma, k) where:

or, equivalently could be defined as DoG(x, y, sigma1, sigma2):

derivation b)

to speed up scale selection, the paper does two things:
iterative blurring of layers
Instead of blurring each layer with the appropriate (k^n)*sigma, which would require a large gaussian (blur) kernel and thus would be computationally expensive, we perform an iterative blurring, taking the image at a particular blur and convolving it with sigma_delta (defined next) to yield the next level. Sigma_delta is effectively much smaller than (k^n)*sigma, therefore the convolution is efficient.
downsampling every octave
Instead of generating the image layers used to make DoG layers at all scales, the paper only generates image layers corresponding to the levels between a given sigma to twice that sigma. Then, to generate the next (and successive) octaves, the existing image layer generated for the current octave corresponding to 2*sigma is downsampled to half its original resolution - a very fast operation, and the process repeated.

derivation c)

Using the gaussian semi-group property we want to find an expression for g(sigma_delta):

g(sigma_delta^2) * g(sigma_cur^2) = g((k*sigma_cur)^2)
since