In simpler terms, how to build a SIFT descriptor?

Question

Programming

Based on this article: https://habrahabr.ru/post/106302/
The first points are clear and simple
1) Gauss passes through the image several times. We save some stages of blurring
2) Subtract the "neighboring" received images
3) We examine each pixel for an extremum together with the points surrounding it of the image itself, as well as the images "above" and "below"
All this is done for several "octaves", i.e. the original image is halved, and so on.
And this is where I fail to understand. What does "approximate the DoG function by a second-order Taylor polynomial" mean, in the sense of what exactly to do with the image in terms of implementation, for example, in C#? I don't understand finding derivatives in pixels.

Answer 1

We'll have to figure out the accompanying math first. As I understand it, the task comes down to writing logic code to determine whether a point in an octave is a key point, so that you can find the key point of an octave.
Or you can try to understand the existing code - see below. here and here .