You have pretty heavy artefacts in the lower left corner, switching between blurred- and sharp images. Also a lot of interleaving - is that from the capture or part of the algorithm?
Could it be just a straight blend between prefiltered images, in which case ~1-2ms? Or summed-area-tables (~3ms?)
I dunno about the interleaving, it shows when brightened and enlarged as I did but not on the original image. It's not a blend between prefiltered images nor SAT, at it's around 1.2ms. Also I said, it supports motion-blur (without changing the timings)...
March 20, 2011 at 8:02 PM
hb said...
ok ill be the guy to ask - what did u do?
if i had to guess i would say u are doing it brute force - doing a gaussian blur, adapting the number of samples based on distance. and skewing the kernel for MB. seems do-able at this res. close?
Won't say right now, but your idea is interesting. Sampling though a full 2d kernel is too slow (the blur buffer res is half-HD720, and the radius I achieve is pretty big, actually it could be pushed even further by spacing the samples a bit)
March 24, 2011 at 5:21 PM
[Image]
Scaled to 300% to ease viewing
(click to enlarge)
It does motion blur as well. Guess how many ms on 360 :)
"DOF Test"
4 Comments -
Very blurry! :)
You have pretty heavy artefacts in the lower left corner, switching between blurred- and sharp images. Also a lot of interleaving - is that from the capture or part of the algorithm?
Could it be just a straight blend between prefiltered images, in which case ~1-2ms? Or summed-area-tables (~3ms?)
March 16, 2011 at 7:41 AM
I dunno about the interleaving, it shows when brightened and enlarged as I did but not on the original image. It's not a blend between prefiltered images nor SAT, at it's around 1.2ms. Also I said, it supports motion-blur (without changing the timings)...
March 20, 2011 at 8:02 PM
ok ill be the guy to ask - what did u do?
if i had to guess i would say u are doing it brute force - doing a gaussian blur, adapting the number of samples based on distance. and skewing the kernel for MB. seems do-able at this res. close?
March 23, 2011 at 1:53 PM
Won't say right now, but your idea is interesting. Sampling though a full 2d kernel is too slow (the blur buffer res is half-HD720, and the radius I achieve is pretty big, actually it could be pushed even further by spacing the samples a bit)
March 24, 2011 at 5:21 PM