Surf learning Note 2

To understand surf, first understand the relationship between the integral image and the Hessian matrix.
In sift, the dog space is first generated, and then the Hessian matrix is calculated in the dog space. In surf, the process of generating a scale space is implicit. For the same integral image, the Hessian matrix of each point on the image is directly calculated using an ever-increasing scale (the theoretical basis for this computation is), and the determining factor of the Hessian matrix is saved, then we can get a scale space (the scale is the functions of Ave ave and interval ).
The problem is
1. What is the basis for calculation of dxx in this way?
2. What is the basis of normalise?
3. What is the basis of lap_sign?
4. determinant = (dxx * dyy-0.81f * dxy); what is the basis?
5. What is m_det?
See the opensurf document. The Hessian determinant value is the product of the feature value. If the two feature values are the same number, then h> 0 indicates the extreme value. Otherwise, no. For more information about the Hessian matrix, see the Wiki matrix.
Hessian matrix is also used in sift.
Also available in Harris corner
Question 1: What We Need To calculate is the Hessian matrix of the Scale Space image. The scale space is obtained by convolution of image I and Gaussian kernels of different scales, and is recorded as l, bay provides a method to quickly calculate the approximate Second-Order Partial direction dxx, dyy, and dxy of L on integral images, so as to quickly calculate the h (x, y, SIGMA) of each point ); the scale space is obtained. The specific octave and internal are divided as follows:
1. The boxfilter corresponding to the initial Sigma = 1.2 is 9*9. There is a formula between the size of Sigma and boxfilter to calculate the approximate Sigma.
2. The minimum size growth is 6, because ....
3. The first scale of the new Ave ave is the last scale of the previous octave.
Question 5 -- we have obtained h (x, y, SIGMA) and saved it in m_det to find interest points.