Progress on landscape makers about neutrallandscapes.jl HOT 21 CLOSED

I can do "random cluster nearest neighbor" :))

from neutrallandscapes.jl.

@mkborregaard I think so - exportasciigrid is something I've done before (readasc as well) so I'll dig it up.

from neutrallandscapes.jl.

Do you want push access to the org @rafaqz ?

from neutrallandscapes.jl.

Sure. Thanks.

How do you feel about an ImageMorphology.jl dep? It has the label and feature distance methods we need for nearest neighbor algs.

from neutrallandscapes.jl.

Given that this is an end-user-facing package rather than a base package I don't see any issues with taking on deps. It's not exactly lightweight though - does NearestNeighbors.jl not do the trick? - let's continue that discussion under #14

from neutrallandscapes.jl.

@mkborregaard just tried pushing a perlin branch but was denied 🥺

Can someone give me push access?

from neutrallandscapes.jl.

Did you accept the membership invite?

from neutrallandscapes.jl.

You didn't, it seems, it's still listed as "pending"

from neutrallandscapes.jl.

Oops sorry, my github message feed is insane these days

from neutrallandscapes.jl.

What's missing in addition to the algorithms?

• classifyarra0
• blendarrays
• blendclusterarray
• meanofcluster
• exportasciigrid [wontfix]

and that's it, right?

from neutrallandscapes.jl.

ascii in #26 -- this is exportasciigrid

from neutrallandscapes.jl.

So the last algo remaining is not the most friendly, because we need a sort of image segmentation code for it to work.

from neutrallandscapes.jl.

is that ndimage.measurements.label?
I worked around that function call in _clusterMean but looking at it I don't think I did it right, so I'll just need to check that code again.

from neutrallandscapes.jl.

It can be solved with a heuristic as a temporary solution. That's nothing more than a constrained label propagation, and this tends to converge in a few iteration. I can try this and see if that performs well.

from neutrallandscapes.jl.

Nice. Yeah I thought it could essentially be done single-pass but maybe I'm not understanding the full complexity of the issue

from neutrallandscapes.jl.

ImageMorphology.jl has this. Maybe just doing that PR to generalise it is less work.

from neutrallandscapes.jl.

So there is very little to it in ImageMorphology... except it uses traits like coords_spatial from ImageCore.jl. Again not even image related, but lots of image related deps.

It looks like we have to abstract everything not image related out of the Images.jl ecosystem at some stage. I think it's not much work technically and it's even their long-term intention, but getting all the PRs through and package registrations would be a lot of work. But ...someone has to do it some day - these methods are in Numpy, not some python image package.

from neutrallandscapes.jl.

Update: I wasn't understanding the full complexity of the issue.
Anyway, pushing on ImageMorphology sounds like a good idea, if that org catches the ball and start running with it some of us could possibly help? Let's see how responsive they are.

from neutrallandscapes.jl.

I could do most of it honestly, they are really small changes. The main issue is getting agreement on moving the traits out of ImageCore.jl and to somewhere without all those depsendencies.

from neutrallandscapes.jl.

I think the ImageMorphology is the right way to go honestly and it sounds like they are amenable to it. Honestly having them split and then Raf's fast code in Stencils would be a really awesome thing for the ecosystem.
Just because I'm stupid I did find it fun to try and implement a label function myself. I think I succeeded (and that it's OK efficient) - I could put that in a PR for now?

function _label(mat)
    rook = ((1, 0), (0, 1))
    m, n = size(mat)
    (m >= 3 && n >= 3) || error("The label algorithm requires the landscape to be at least 3 cells in each direction")
    
    # initialize objects and fill corners of ret
    ret = zeros(Int, m, n)
    clusters = Dict{Int, Int}()
    label = 0

    # run through the matrix and make clusters
    for j in axes(mat, 2), i in axes(mat, 1)
        same = [i - n[1] > 0 && j - n[2] > 0 && mat[i - n[1], j - n[2]] == mat[i, j] for n in rook]
        if count(same) == 0
            ret[i, j] = label += 1
            clusters[label] = label
        else
            vals = [ret[i - n[1], j - n[2]] for n in rook[same]]
            mi = minimum(vals)
            for v in vals
                clusters[v] = mi
            end
            ret[i, j] = mi
        end
    end
    
    # merge adjacent clusters with same value
    ncl = 0
    for i in eachindex(ret)
        ret[i] = _getval(clusters, ret[i])
        ncl = max(ncl, ret[i])
    end
    ncl, ret
end

function _getval(clusters, val)
    while true
        val2 = clusters[val]
        val2 == val && return val
        val = val2
    end
end

An example use

a = rand(DistanceGradient(rand(1:100, 3)), 10, 10)
b = classify!(copy(a), ones(4))
nc, c = _label(b)
plot(heatmap(a), heatmap(b), heatmap(c, fillcolor = :picasso), size = (1400, 400), layout = (1,3))

Two example runs

from neutrallandscapes.jl.

that looks like it works - let's PR this and if the ImageMorpho split happens we'll replace it?

from neutrallandscapes.jl.