It works. On everything.
It's not guaranteed to always do better than the extant techniques, but on most metrics, almost all the time, it is noticeably better. (There are two methods that can in theory beat it by a fair margin, but they require a priori knowledge that doesn't always exist in practice. And they fail badly if you guess wrong.) I also don't have code to compare it with the crazy
complicated method that inspired me, but since that's just a piecewise approximation of
what I'm doing, I'm pretty confident I've got it beat.
And the actual method itself, once you've got everything set up? 7 lines of code.
I can haz brain back nao, plz?