The goal is to solve hard-to-get-rid-of noise with really bright
specular highlights which are out of focus by doing some automatic
blur on glossy roughness.
The idea is simple: calculate PDF of a camera ray and use it as initial
value for filter glossy.
Implementation is mainly straighforward: approximate blur size and use
it's ratio to initial aperture size as PDF (kind of like geometric
probability). While this seems like it kind of works, there must be
some easier to control approach because currently this requires some
manual tweaking of extra argument to avoid too much blur and to make
it to work nicely with regular filter glossy.
Seems we'll need extra argument anyway to preserve compatibility, but
some equations we can modify quite easily.
Here is some artificial file to show things in action (NOTE: this
feature doesn't necesserily help in this setup. Main purpose is to
demonstrate how it all works).
Without Filter Glossy DoF:
With Filter Glossy DoF of 0.5:
The idea of improvement could be to check how bidirectional path
traces are estimating PDF for such cases.