Hemispheric Sampling

I'm curious about the hemispheric sampling referenced in this presentation.

http://www.radiance-online.org/community/workshops/2011-berkeley-ca...

If I'm interested in spherical sampling of a space, how would that work when you want to sample something in open space rather than on a surface? I noticed that the results are highly dependent on normal of the surface used to generate the sample points and mesh. This poses a problem when looking to get a good idea of how light behaves in an open space like an atrium sampling in section (see images). I'll create some additional examples. 

Would you generate results for the two hemispheres and combine results additively?