Distinctive Image Features from Scale-Invariant Keypoints
Distinctive Image Features from Scale-Invariant Keypoints Also known as Scale-invariant feature transform (SIFT) or SIFT features.
Author: David G. Lowe
Affiliation: University of British Columbia
Algorithm
Scale-space extrema detection
\(L(x,y,\sigma)\) is the scale space of an image. This is generated by the convolution of a Gaussian with an input image: \[ \begin{align} L(x,y,\sigma) = G(x, y, \sigma) * I(x,y) \end{align} \] where \(G(x,y,\sigma) = \frac{1}{2 \pi \sigma^2}\exp{-(x^2+y^2)/(2\sigma^2)}\).
This is used to compute the difference-of-Gaussian (DoG) for two scales separated by \(k\): \[ \begin{align} D(x,y,\sigma) &= (G(x,y,k\sigma) - G(x,y,\sigma)) * I(x,y)\nonumber \\ &= L(x,y,k\sigma) - L(x,y,\sigma) \end{align} \]
For your image, you create multiple octaves by down-sampling the image by 2.
Within each octave, you have multiple scales with varying \(\sigma\).
E.g. if you have 5 scales per octave, then you will have 4 DoG images.
To find local min/max, compare each pixel in each DoG image to its 8 neighboring pixels within the image and 9 neighboring pixels in each neighboring scale. In total, it will be the min/max among (8+9+9=26) neighboring pixels. This is only done on the intermediate DoG images (i.e. excluding the first and last).
In his paper, Lowe uses \(k=\sqrt{2}=2^{1/s}\) with \(s=2\). This means generating \(s+3=5\) Gaussian blurred images for each octave.