[FEATURE] make VND.bw a 3 index array in preparation for parametric

[dehann]

Probably a DFG issue, but starting it here. Need is to store off-diagonal covariances while using the same VND data structure.

[dehann]

cc @Affie

[dehann]

we should look at what the performance characteristics would be.

[Affie]

I did a quick comparison test.
Array
0.026597 seconds (8 allocations: 22.888 MiB, 55.53% gc time)
Tuple
0.028543 seconds (8 allocations: 22.888 MiB, 29.23% gc time)
Struct
0.015862 seconds (8 allocations: 22.888 MiB, 23.56% gc time)
Mutable Struct
0.036606 seconds (1.00 M allocations: 38.147 MiB, 37.92% gc time)
I will look at the options when I have a bit of time

[Affie]

Starting of we are only working with one Gaussian mode. Therefore the covariance can fit in VND.bw,
I do not believe that an array is the easiest way forward. I propose that we wait with this changes that we might or might not need in the future. For instance, having a typed array might fit in a lot better when working with manifolds such as SE3. We can consider changing to types with a get/set index defined on them to make it easier to transition from the current usage.

DF, Agree lets think about doing a Vector of group elements for val and then supporting information like bw. Maybe these need to be nested types, although it's a little difficult to know precisely which type will work long term. Lets hold with bw::Array{<:Real,2} for now then.

[dehann]

Hi thanks for taking a look! Wait on accessors -- got it; and TypedArrays ... havent used or even heard of them myself, will go read up some more, thanks again.

[Affie]

What I meant was; for example an array of type SE3 Vector{SE3} and not Array{Float64,2}

[dehann]

see edit here: #465 (comment)