Re: Introducing generation of non-fixed size filters/kernels

From: Mateusz Loskot (mateusz_at_[hidden])
Date: 2019-07-21 21:14:54

On Sun, 21 Jul 2019 at 23:07, Olzhas Zhumabek
<anonymous.from.applecity_at_[hidden]> wrote:
>
> Mateusz mentioned that I could use kernel_1d types,

I'd say, I've asked question for my own better understanding.
For complete picture, I'll copy my questions from [1]

I'm interested to understand what's the motivation of using gray32f_view_t as
a kernel type instead of Numeric's types: kernel_1d and kernel_1d _fixed.
If the proposed generators use the kernel_1d* types, then produced kernels
are directly usable with the Numeric's convolve functions.

AFAIK, Stefan is going to move Numeric extension features to the core
what will make the kernel types available.

[1] https://github.com/boostorg/gil/pull/342#pullrequestreview-264528946

> but since they don't
> share common root in project tree, I guess I should perform migration from
> image_view to kernel_1d in a separate PR when numeric extensions become
> part of core. What do you think?

I wouldn't say you should and you probably have higher priorities for now.
I just noticed that
- we have types for kernels
- we have functions that take the existing types of kernels
- it would be useful if any new kernel-specific features integrate
  with the existing ones.

> By the way, currently Gaussian kernel generation tests tolerate 5%
> inaccuracy wrt to expected values from wikipedia page. I didn't try lower,
> should I attempt to find the smallest value?

>From my point, if it not a high priority for your own project,
I would open an issue to investigate that in future (an issue
will help to  not to forget about it), and postpone it for now.

Best regards.
-- 
Mateusz Loskot, http://mateusz.loskot.net