Add smoothing of contour plots using Gaussian filtering

[Cornelius-G]

Allows smoothing of contour plots using Gaussian filtering (via ImageFiltering.jl).

Without smoothing:
p = plot(samples, (:a, :c), st=:smallest_intervals_contour, bins=100, smoothing=0)

With smoothing:
p = plot(samples, (:a, :c), st=:smallest_intervals_contour, bins=100, smoothing=1.5)

Also possible to smooth the dimensions separately:
p = plot(samples, (:a, :c), st=:smallest_intervals_contour, bins=100, smoothing=(0, 1.5))

(Also replaced all remaining "local modes" by "marginal modes")

[codecov-io]

Codecov Report

Merging #268 (9db5d58) into master (a24f3d6) will increase coverage by 2.68%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##           master     #268      +/-   ##
==========================================
+ Coverage   55.60%   58.28%   +2.68%     
==========================================
  Files         109      109              
  Lines        5257     5550     +293     
==========================================
+ Hits         2923     3235     +312     
+ Misses       2334     2315      -19     

Impacted Files	Coverage Δ
src/BAT.jl	85.71% <ø> (+2.38%)	⬆️
src/algotypes/mode_estimator.jl	37.50% <ø> (+8.92%)	⬆️
src/plotting/recipes_MarginalDist_2D.jl	0.00% <0.00%> (ø)
src/plotting/recipes_samples_1D.jl	0.00% <0.00%> (ø)
src/plotting/recipes_samples_2D.jl	0.00% <0.00%> (ø)
src/densities/abstract_density.jl	97.29% <0.00%> (-1.24%)	⬇️
src/utils/coord_utils.jl	100.00% <0.00%> (ø)
src/statistics/autocor.jl	100.00% <0.00%> (ø)
src/distributions/funnel.jl	100.00% <0.00%> (ø)
src/densities/const_density.jl	100.00% <0.00%> (ø)
... and 69 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update a24f3d6...9db5d58. Read the comment docs.

[oschulz]

This is very nice - only, ImageFiltering would be quite a heavy additional dependency for BAT, it pulls in several other imaging-related packages that we don't need at all. Can we find a similar functionality/algorithms in a more statistics-oriented Julia package?

[oschulz]

We already depend on KernelDensity.jl, could we maybe do this via BivariateKDE?

[Cornelius-G]

We already depend on KernelDensity.jl, could we maybe do this via BivariateKDE?

This was actually also my first idea. However, the things I tried with KDEs (and also with grid interpolations) did not really improve the contours.
I'll see if I can find some more light-weight filtering algorithms.

[Cornelius-G]

So, I removed the dependency on ImageFiltering.jl and added a file convolution_utils.jl that contains very basic code for a convolution (including padding) and Gaussian kernels.
The implementation is really simplistic and quickly put together from the code here and here, but seems to work.
Unfortunately, I currently don't have the time to make this much nicer, but would like to quickly add this feature. Maybe someone else would like to do a bit of code reduction and profiling.

However, there are efforts for separating the filtering stuff from the image-related code in ImageFiltering.jl being currently discussed. So, we might want to keep an eye on that and switch back to a more sophisticated implementation later.

[Cornelius-G]

@oschulz GitHub Actions had some issues yesterday evening, so the CI failed or is still queued. I don't see the button to restart CI. Can you run this again somehow?

[oschulz]

Yes, I was wondering ... restarting it now.

[oschulz]

Nice!