Check differential factors use correct approxDeconv result

[dehann]

• approxDeconv recently changed to return tangents, previously was coordinates.
• differential factors based on MKD should be built with manifold points.

RoME test on IIF913 produced this error
(stable/good on left -- current master/bad on right):

[Affie]

It sounds like you are on the right track, I previously tracked the issue to approxDeconv and fixed it here #1380, but I didn't have time to dig deeper.

Just note that the sampleTangent function should be able to handle the correct sampling already. So if MKD is built correctly it should work.

[dehann]

I think the bug is happening here:

IncrementalInference.jl/src/TreeMessageUtils.jl

Lines 314 to 316 in 145774f

	pts, = approxDeconv(tfg, afc.label, solveKey) # solveFactorMeasurements
	# @show sft
	newBel = manikde!(pts, sft)

[Affie]

Here is the test code I have so already, but it's focused on approxDeconv:

IncrementalInference.jl/test/testSpecialEuclidean2Mani.jl

Lines 189 to 232 in 145774f

	@testset "test deconv on <:AbstractManifoldMinimize" begin
	##
	fg = initfg()
	getSolverParams(fg).useMsgLikelihoods = true
	addVariable!(fg, :x0, SpecialEuclidean2)
	addVariable!(fg, :x1, SpecialEuclidean2)
	mp = ManifoldPrior(SpecialEuclidean(2), ProductRepr(@MVector([10.0,10.0]), @MMatrix([-1.0 0.0; 0.0 -1.0])), MvNormal([0.1, 0.1, 0.01]))
	p = addFactor!(fg, [:x0], mp)
	doautoinit!(fg,:x0)
	addFactor!(fg, [:x0;:x1], ManifoldFactorSE2(MvNormal([10.0,0,0.1], diagm([0.5,0.5,0.05].^2))))
	initAll!(fg)
	# now check deconv
	pred, meas = approxDeconv(fg, :x0x1f1)
	@test mmd(SpecialEuclidean(2), pred, meas) < 1e-1
	p_t = map(x->x.parts[1], pred)
	m_t = map(x->x.parts[1], meas)
	p_θ = map(x->x.parts[2][2], pred)
	m_θ = map(x->x.parts[2][2], meas)
	@test isapprox(mean(p_θ), 0.1, atol=0.02)
	@test isapprox(std(p_θ), 0.05, atol=0.02)
	@test isapprox(mean(p_t), [10,0], atol=0.2)
	@test isapprox(std(p_t), [0.5,0.5], atol=0.2)
	@test isapprox(mean(p_θ), mean(m_θ), atol=0.02)
	@test isapprox(std(p_θ), std(m_θ), atol=0.02)
	@test isapprox(mean(p_t), mean(m_t), atol=0.2)
	@test isapprox(std(p_t), std(m_t), atol=0.2)
	end

[Affie]

I think the bug is happening here

Yes, that needs to be converted to points. The current assumption is tangent on lie group so i think we can just use exp map at identity for now?

pred_X, _ = approxDeconv(tfg, afc.label, solveKey)
pts = exp(M, ident, pred_X)

[dehann]

Cool thanks, yes let me try that...

[dehann]

and getting the following error (stable/good on left -- current master/bad on right):