Fixing bug in Harris/Hessian examples

From: Olzhas Zhumabek (anonymous.from.applecity_at_[hidden])
Date: 2019-09-11 23:25:17

Hi,

As discussed on Gitter, I found a problem with overflow in the examples, as
they use unsigned storage for gradients ... I would like to fix that by
following:

1. Create a generate_x/yscharr_kernel and generate_x/ysobel_kernel
functions, with a degree of differentiation as argument (defaulted to one)

There are two widespread kernels to compute gradients, and I believe it
should also be possible to get a kernel that would differentiate multiple
times with one convolution.

2. Write tests for 1

3. Create adaptor interface for convolution_2d using image_view, to solve
circular test dependency problem (by copying the image_view contents into
kernel_2d)

Scharr kernel is not separable [[-1, 0, 1], [-1, 0, 1], [-1, 0, 1]] as far
as I know, thus 2d convolution is required

4. Rewrite examples to use the new functionality

What do you think?

Olzhas