Feature Extraction

Feature Extraction

HOG – histogram of gradients

corner detection:

compare image intensity within a window with a neighbourhood of windows

change is happening in multiple directions

Harris corner:

uses moving window

significant change in all directions

weight function in window (usually use gaussian)

compute the difference using small shifts around the window

expensive to compute naively O(window_width^2 * shift_range^2 *image_width^2)

want to compute value of function around current location

use taylor series expansion (approximate function around location)

eigen value decomposition to get the eigenvalues of a window (smaller values will mean that there isn’t much happening. Corners will have peaks at where the edges are and will have high eigenvalues)

plotting eigenvalues:

edge: one lambda is larger than the other

corner: both lambdas are large and fairly similar

Steps:

compute image gradient in x and y

scale selection

Laplacian of Gaussian intereest point

vary sigma for laplacian function for small image

compare pixel value for neighbourhood of pixels and for pixels with higher and lower laplacian

strongest pixel is selected within (x,y,s)

find interest point at location and scale

Difference of Gaussian can be used as an approximation of Laplacian of Gaussian

compute difference for multiple sigma values

compute the same for multiple resolutions

LoD is not separable, while DoG is separable

Hessian interest point

use second derivative instead of first derivative

compute ratio between trace and determinant of Hessian matrix and choose key points which pass a threshold (10)