Yeah, looking at it again, it just appears to be "Create shadow map like structure, create frustum for each primitive in the shadow map, link onscreen pixels to shadow map texels (list of onscreen pixels as covered by which shadow map texel) then for each pixel, go to shadow map texel it's linked to, then within that texel test against each frustum you constructed to see whether that point (pixel I think) is within the frustum. It's a lot like shadow volumes, but unfortunately seems to involve atomics (mentioned as part of the irregular z-buffer, which creates the linked list of onscreen pixels to shadow map texels) as well as creating frustums and then frustum tests.
None of it by itself is necessarily expensive, but is overall a lot more expensive than just doing a normal shadow map, let alone the virtual shadow map I originally linked to which can be faster than a normal shadow naive cascaded shadow map. Virtual shadow maps will also come close to having a 1 to 1 match to screen resolution, and so will have close to the same quality as this anyway while being a lot more efficient. The only other thing they're doing is comparing unfiltered shadow maps without SMAA applied to their results (With SMAA applied, which has nothing specifically to do with this, cough cough) and frankly no one just uses unfiltered shadow maps and I had to look at the zoomed in portion of their comparison to appreciate the difference anyway.