A while ago I wrote about high quality shadow filtering, and in particular about how to approximate a Gaussian using bilinear taps. Today I just noticed the following in the third edition of Real-Time Rendering:
Alternatively, the Gaussian could be approximated by using a bilinear interpolated sample for every four texels, finding the offset that gives the closest approximation to the ideal weights.
So, I guess I was not saying anything new after all.
4 Comments
Most of the time everything turns out to be already invented. Usually in the 70’s by Blinn and friends… Ok, I’m slightly exaggerating here :)
Oh, but you were saying something new. We were saying “this is possible”, and you were saying “this is how it’s done”, which is much more useful :-)
I agree with Naty! I’m the guy that wrote that line for RTR3 (or maybe it was Naty revising my text). It was just one of those things that seemed worth noting, a pretty natural extension of doing this for the box filter. In practice almost everyone does a one-dimensional two-pass operation for a Gaussian in screen space, so I didn’t think more about it.
What I like about your article is that you show an actual use for a single-pass Gaussian, describe how you actually compute the coefficients, plus use Gather4. And, like Aras says, lots of stuff gets invented, but gets buried on page 471 of some tome or never gets published at all. Sometimes it’s almost better to not know it’s already written up somewhere else, as you might improve upon the published method vs. just implementing it.
I published a tiny paper in JGT about a simple trick I came up with that *nobody* I knew had ever heard of… and one of the anonymous reviewers was all ‘man, this is a totally standard thing and has even published before, for example, ok, here it is in this sample code in this book, although they don’t call any attention to it’. But then he asked around his workplace and nobody else knew about it, so he agreed it was worth publishing anyway.
( http://www.nothings.org/computer/edgeenum.html )
One Trackback
[...] post about Gaussian filtering of shadow maps, though his observation is (as with almost everything) previously known. This pair of forum post by Jonathan “LoneSock” Dummer discuss variants of VSM. Brian [...]