amonakov, 3 months ago (edited 3 months ago)

@demofox Yes! Young and Van Vliet described a fast approximation for Gaussian filters via IIR. It is used in RawTherapee and uncovered a very peculiar bug in GCC!

(Young – Van Vliet gives separate IIRs for each dimension, so at least one of them is very SIMD-friendly)

NOTimothyLottes, 3 months ago

deleted_by_author

demofox, 3 months ago

@NOTimothyLottes that's true about temporal! Bart has an article on spatial looks like, and he said the same as you.

BartWronski, 3 months ago

@demofox yes: https://www.inf.ufrgs.br/~eslgastal/DomainTransform/ we used it on mobile. Not worth it, complete pain in the ass to do it efficiently, code hard to read without a PhD in numerical methods.

BartWronski, 3 months ago

@demofox also https://bartwronski.com/2022/09/05/gradient-descent-optimized-recursive-filters-for-deconvolution-deblurring/ :)
(You need to run 2 passes, or 4 - total pain)

demofox, 3 months ago

@BartWronski thanks for satisfying my curiosity. That's cool stuff

demofox, 3 months ago

@BartWronski I finally got a moment to read through the whole thing. Super neat stuff.
Near the beginning when you algebraically turned a 2 tap box filter into its IIR inverse was a crazy cool insight.

BartWronski, 3 months ago

@demofox What's fascinating is that most convolution filters have very nasty and infinite deconvolution formulas.
This is why IIR for this specific task actually makes sense!
If you are interested in why is that, I have another post on deconvolution with inverting matrices and even more algebra! https://bartwronski.com/2022/05/26/removing-blur-from-images-deconvolution-and-using-optimized-simple-filters/

I know you are a cool always-learning nerd, so I recommend it! :) 1/2