Following up on our Slack discussion: since the API for distributions is still taking shape and capacity is currently limited, I'm opening this issue to track in the long term.
Hi everyone. In JuliaRobotics, we are looking to standardize how we represent and store distributions on manifolds. I see ManifoldDistributions.jl is currently mostly a skeleton, so I'm hoping to get some advice on the intended API pattern.
For our immediate use cases, we primarily rely on "concentrated Gaussians" (by which I mean mapping a Gaussian from the tangent space $T_p \mathcal{M}$ via the exponential map, but concentrated enough that we currently ignore the manifold volume correction for normalization). We also build KDE-like distributions using these as kernels.
I've seen the basic distribution primitives currently inside Manifolds.jl, and I also noticed in the recent GeometricKalman.jl package that you sidestep the Distributions.jl interface entirely by keeping the manifold, the retraction, and a standard MvNormal completely separate.
If we want to formalize this into proper structs (and eventually support true volume corrections), what is the recommended pattern? My goal is to align with whatever pattern ManifoldDistributions.jl has planned so we can easily upstream this work later if needed.
Related to #3
Following up on our Slack discussion: since the API for distributions is still taking shape and capacity is currently limited, I'm opening this issue to track in the long term.
Hi everyone. In
JuliaRobotics, we are looking to standardize how we represent and store distributions on manifolds. I seeManifoldDistributions.jlis currently mostly a skeleton, so I'm hoping to get some advice on the intended API pattern.For our immediate use cases, we primarily rely on "concentrated Gaussians" (by which I mean mapping a Gaussian from the tangent space$T_p \mathcal{M}$ via the exponential map, but concentrated enough that we currently ignore the manifold volume correction for normalization). We also build KDE-like distributions using these as kernels.
I've seen the basic distribution primitives currently inside
Manifolds.jl, and I also noticed in the recentGeometricKalman.jlpackage that you sidestep theDistributions.jlinterface entirely by keeping the manifold, the retraction, and a standardMvNormalcompletely separate.If we want to formalize this into proper structs (and eventually support true volume corrections), what is the recommended pattern? My goal is to align with whatever pattern
ManifoldDistributions.jlhas planned so we can easily upstream this work later if needed.Related to #3