RED Rushes Resample Filters Comparison

[Dj Joofa]

I once tried a Kaiser with a lot of lobes and, while the MatLab graph was pretty, the visual results were not.

Hmm. What were you doing? Kaiser window is a very optimal window in energy terms (prolate spheroidal functions). BTW, I once met Kaiser (or was that his son?).

I think engineering needs to be more bottom up, about what looks good (and trying to figure out how our visual system works), instead of sticking to particular models about how it "should" work.

Such models fall into the domain of what is called perceptual organization and grouping -- bottom's up processing by human vision into meaningful useful information. There is enough indication that mid to high level stages of human vision make use of perceptual organization, however, that is mostly for extraction of edges for accumulating hierarchical meaningful structures. Frequency manipulation is not handled at this stage of vision. Frequency processing is done at the early stages.

Please see the standard references on this topic such as books by Boynton (not Poynton) and David Lowe, etc.

2- If I wasn't lazy I'd try and figure out something in Shake. Do sinc downsampling, except make a matte to catch the edges and use a different algorithm for edges to avoid ringing artifacts/halos on edges.
(Well, that's the kind of adaptiveness I had in mind anyways!)

Several authors have proposes solutions similar to what you are suggesting above. Please pm me your email address and I should send you some information on this area.

 

[Dj Joofa]

what i think there is a need to understand is how different algorithms are going to translate into visual results

Hi DArcyFD, I have put a comparison of various popular image resampling algorithms on a few standard images on my website in the two links below. Just click them!

Comparison is between Gauss, Bilinear, Nearest Neighbor, Mitchel, BiCubic, Hermite, Hanning, Lanczos. They are extracted from a paper, therefore the "Our Algorithm" tag refers to an approach by the authors where they did something similar to what Glenn was suggesting of using edge information while resampling.

Resampling Comparison # 1
Resampling Comparison # 2

EDIT: The files will be back shortly!

Unfortunately, the image quality is small and not too good. If you pm me your email, I can send you some more information.

 

[D'Arcy Foley-Dawson]

....can't wait to see these pics!....

 

[Dj Joofa]

....can't wait to see these pics!....

Hi DArcyFD, I am providing an alternate link to a comparison done by some other people. I have delayed posting the pictures on my website, as I don't want to get embroiled in copyright concerns over posting stuff publicly.

Before you go to the following link, please read the following:

(1) To me it appears that the authors in the following link are also under the erroneous impression, which is a common mistake several people make, even to this date, that the use of B-splines blurs/degrades image interpolation quality. Several people have claimed so in prestigious IEEE prublications and image processing textbook -- even the 1978 edition of one of the most famous textbooks in image processing, Digital Image Processing by W. K. Pratt, also did so. This is generally caused by an incorrect implementation where pixel values are used directly for B-Spline co-efficients.

(2) Such confusion is partly caused by difference in terminology between engineers and mathematicians in the field of approximation theory, what engineers call bandpass is meant to be entire functions of a certain type in mathematics literature. (And we will not talk about physicists, as their notation and way of explaining things is utterly pathetic! Mathematicians and engineers are fine.)

(3) In case of downsampling, experiments should be done with interpolant parameters, etc. In the following link, I think they did not explore a whole lot, and perhaps that is why some of picture with spline interpolation look too blocky.

Click here.

 

[Richard Starkey]

Resample Filter Comparison

Resample Filter Comparison

So I get enough free info from these sites that it is now time to share.
I have done the boring job of trying the same clip on all the Resample Filters in REDrushes and here are my results:
One clip, 667 frames, not 4K but that new RED anamorphic aspect ratio (another discussion)
Processed on brand new, out-the-box iMac Quad 27" (converting to Quicktime PRORES SQ)

Quarter-RES
Quadratic - 1 min 14 sec
Bell - 1:14
Mitchell - 1:16
Gauss - 1:17
Cubic - 1:17
CatmulRom - 1:18
Lanzcos - 1:28
Widegauss - 1:38

Half-RES
CatmulRom - 2:25
Bell - 2:15

The half- res is pretty much as you would expect- twice as good as the quarter-res and almost twice as long.
Even though I am only offlining at this res I can/t stand those jaggesy lines that you get from 1/4 res- will be too embarrassed to show the DOP that s**t.

The quality: tricky one but I think the sharp ones look better than the soft ones if you go 1/4 res. However, if you can be bothered going half-res then Bell looks pretty sweet and not too soft.
My picks are the two that I tried at half-res. Either one. In fact I can't even remember what I left running while I popped out for dinner!

 

[David Battistella]

But this is all on the "old" color science , correct?

david