I’m not that familiar with previous work in the area, so I don’t know how it stacks up against e.g. Gigavoxels; if anyone has any insight please let me know.  One interesting detail from the paper was the way they enhance each cubical voxel with a pair of bounding planes, which they call contours.

They show that the data structure enables casting rays significantly faster than against a triangle model of equivalent complexity.  Unfortunately, it appears to still be slower than rendering the triangles the “old-fashioned way”.  The technique also requires post-filtering since it does not allow for filtering of color and normal information (which is effectively point-sampled).  Finally, building the data structure takes a fair bit of time, making this not particularly well-suited for dynamic scenes.