Returns HOG block descriptors computed for the whole image

descs = hog.compute(im)
descs = hog.compute(im, 'OptionName', optionValue, ...)

Input

• im 8-bit 1- or 3-channel source image where HOG features will be calculated.

Output

• descs Row vectors of HOG descriptors, with the number of columns equal to hog.getDescriptorSize(). Matrix of type single.

Options

• WinStride Window stride [w,h]. It must be a multiple of block stride. Not set by default in which case it uses CellSize.
• Padding Optional padding [w,h]. default [0,0]
• Locations cell array of 2D points {[x,y],...} at which descriptors are computed. Not set by default (in which case descriptors are computed for the whole image with a sliding window).

Computes HOG descriptors of given image.

In case of "dense" descriptors (i.e Locations is not set), the number of rows is equal to the number of sliding windows over the image. Assuming zero padding, this is computed in the following way:

[h,w,~] = size(im);
windows_per_img = ([w,h] - hog.WinSize) ./ WinStride + 1
num_windows = prod(windows_per_img)
% num_windows = numel(hog.WinSize(1):WinStride(1):w) * ...
%               numel(hog.WinSize(2):WinStride(2):h)

The windows cover the image in a top-to-bottom left-to-right order.

In case of "sparse" descriptors (i.e Locations is set), the number of rows is equal to the number of locations specified.

The function is mainly used to learn the classifier.

The computed feature vectors are compatible with the INRIA Object Detection and Localization Toolkit OLT.