[RFC] Redesign the Drawable type

[johnnychen94]

The current ImageDraw design has significantly restricted its possibility to the 2D case, and that's not good in general (even though it's the most common case).

What I have in mind is:

abstract type Drawable end

struct Point{C<:Colorant, N} <: Drawable
    pos::CartesianIndex{N}
    color::C
    thickness::Int
end

struct Polygon{C<:Colorant, N, T<:CartesianIndex} <: Drawable
    vertices::NTuple{N, T}
    border_color::C
    border_width::Int
    fill_color::C
    fill::Bool # whether we need to fill the region with color C
end

struct Layers{A:: PriorityQueue} <: Drawable
    objects::A
end

Basically, each drawable type consists of the whole information on how it can be drawn.

Layers is a simple wrapper on DataStructures.PriortyQueue that controls how individual objects are drawn in one draw call.

Point is not serving as the index, but instead as a real physical "point" with color and thickness information.

This is a draft type design, we could definitely add more information into each drawable struct.

cc: @jiviteshjain @mgautam98

[johnnychen94]

This is so breaking a design that we may want to start a new repo to do this, a more appropriate place for this is in JuliaGraphics because Image is just one of many drawable types:

struct Image{A<:AbstractArray} <: Drawable
    # the axes of image contain all the positional information
    data::A
end

I could initialize it this weekend, in the meantime, we could still continue the devs here.

Edit:

I'm becoming rather ambitious that I'm now considering what graphic processing should be with image processing as one special case of it.

[kimikage]

https://xkcd.com/927/ 😄

Of course, I don't hate reinventing the wheel. My concerns are:

• Graphics.jl doesn't do much for abstraction and is a bottleneck on the dependency chain (cf. CompatHelper: bump compat for "Colors" to "0.12" ImageCore.jl#124).
• GeometryTypes.jl was discontinued, but it is still used in many packages.

[johnnychen94]

Yes, this should start as an experiment and I'm not a graphic expert.

I really don't like the hard-coded x, y, width, height in those packages, given that we have much clearer CartesianIndex. Also, it's not a good experience when I get lost in the countless keyword arguments and names.

It should just start as an experimental package, and if it works well, I'll get ImageDraw deprecated.

[tp2750]

This sounds very interesting @johnnychen94 . Did you start up a new repo?