From the paper: "We use the rasterizer as a potentially-visible set optimization to iterate only over pixels for which rays might intersect a voxel, and then execute a small ray tracer in the pixel shader. That is, we “splat” billboards that give coarse visibility and compute exact visibility in a pixel shader. This works for any pinhole perspective projection, including eye rays and shadow rays, so we use it for the shadow map rendering pass as well"

Neat - I used a similar technique [1] in the last iteration of Voxel Quest, but I only rasterized single points/pixels then raytraced them into filled cubes (was not efficient other than on fill rate). However I did not invent the technique (I'm not sure who did, but I first got the idea from talking to Florian Boesch [2])

[1] https://twitter.com/gavanw/status/717265068086308865 [2] http://codeflow.org