Multi-Label Spherical Dilation

[william-silversmith]

I think it may be possible to configure the marching parabolas EDT algorithm to track the nearest label. This is an easier problem than computing the voronoi diagram for the feature transform. It would be faster and lower memory too.

[MariusCausemann]

Hi @william-silversmith ,
I'm interested in the multilabel spherical dilation. Is this something you are considering to implement in the near future?
Best,
Marius

PS: thanks for fastmorph!

[william-silversmith]

Hi Marius,

For spherical dilate, how are you imagining handling the boundaries between touching objects? No change (i.e. background only) or something like mode? For the former, it may be a bit easier to put something together.

[MariusCausemann]

I think background only would be great!

[william-silversmith]

Hi @MariusCausemann I've added support for spherical dilation, opening, and closing via scipy in 1.7.0. Hope that helps!

If I wrote my own feature_map in edt, it would probably be a lot faster but this is a great stop-gap.

[MariusCausemann]

Hi @william-silversmith ,
thanks a lot for your efforts! I realized that I probably won't get sufficient performance for small radii with an edt based approach, and wrote my own little package with a pseudo-spherical structuring element (alternating box and diamond) with Numba (https://github.com/MariusCausemann/nbmorph).

BTW, I had some success with sorting networks for the mode computation, it might give you a nice speedup over std::sort, too.

[william-silversmith]

That's a good tip on the sorting networks. I've used them to great effect in other packages, usually about doubling sort performance. I'll make that change!