software uprez vs. print at 200%

[Schewe]

I was on a panel on raw processing about a year ago, and so was Eric Chan. At the break, I congratulated him on the big improvements in resizing in Lightroom 4. He said that the algorithms were similar; the big change was that, rather than doing the calcs in the standard Lr working space, which has a gamma of one, they were doing the resizing in a gamma encoded space which is more perceptually uniform. We didn't discuss the gamma, but it's probably either the ProPhotoRGB gamma of 1.8, or 2.2.

Well, you've discovered Eric's secret :~)

In an attempt to improve certain aspects of the final resampling and sharpening, it was determined that sharpening in linear space was suboptimal. I'm not sure when Eric made the change, but there was a certain point where some breakthroughs in Eric's testing lead to the improvements in LR's resampling and sharpening for output.

Not for nothing, if anybody can advocate additional improvements, I'm sure Eric would like to hear about it :~)

[Bart_van_der_Wolf]

In an attempt to improve certain aspects of the final resampling and sharpening, it was determined that sharpening in linear space was suboptimal.

Hi Jeff,

I'm afraid that's not universally true, but I assume Eric knows that but also has to deal with complexity trade-offs of the current sharpening dialog (it could use a serious overhaul). Deconvolution makes more sense in a linear gamma space, edge enhancement makes more sense in a gamma pre-compensated space (to reduce differences in acutance on the light side and the dark side of the edges/lines), and it would benefit from a possibility to target lighter, medium, and darker tones because their contrast is different in non-linear gamma space.

I'm not sure when Eric made the change, but there was a certain point where some breakthroughs in Eric's testing lead to the improvements in LR's resampling and sharpening for output.

He probably, amongst others, paid attention to the discussions on the ImageMagick board and website where it was concluded quite some time ago that downsampling would benefit from linear gamma space filtering, and upsampling would perceptually benefit from doing that in non-linear gamma space.

Professor Nicolas Robidoux also did extensive trials of various methods used by ImageMagick, including temporarily (sigmoidal) contrast adjusted upsampling, especially on the more advanced Elliptical Weighted Averaging (EWA) filtered approaches but also on the tensor based variants of Lanczos windowing.

Cheers,
Bart

[laughingbear]

Hi Bart,

On a side note, I wonder why Adobe had not sent you an invitation and hired you like Eric many years ago already! With your Kodak background and overall expertise, it certainly would be a win-win. Then again, who knows, perhaps you would have declined the invite. 

On this occasion, and seriously, many thanks for years of ruthless expertise sharing.

Hope all is well with you.

Best
Georg

[Bart_van_der_Wolf]

Hi Bart,

On a side note, I wonder why Adobe had not sent you an invitation and hired you like Eric many years ago already! With your Kodak background and overall expertise, it certainly would be a win-win. Then again, who knows, perhaps you would have declined the invite. 

On this occasion, and seriously, many thanks for years of ruthless expertise sharing.

Hope all is well with you.

Hi Georg,

Thanks for the kind words. I'm fine, hope you are fine as well.

I suppose Adobe (and some others), prefer to work with local (USA) folks.

I'm always looking for new developments in all sorts of fields (lots of interesting research going on), and tend to combine the best practices and high potentials from those various fields with common sense and pragmatism, sometimes with surprising results when applied to a specific (e.g. photographic) issue. I know when good enough is good enough, but that doesn't mean one should stop to try raising the bar and differentiate from competition.

Aim low, and you'll shoot something low, aim higher and you'll shoot something higher ... (metaphorically speaking, I do not endorse the use of guns )

Cheers,
Bart

[laughingbear]

I know when good enough is good enough….

That in itself is an art that "develops" with experience only.

I'll send you a PM with my updated contact details. Would love to have your thoughts on the recent Sony CMOS development and inevitable? (Software solution possible perhaps, I am not sure.) channel crosstalk.

[NicolasRobidoux]

...
He probably, amongst others, paid attention to the discussions on the ImageMagick board and website where it was concluded quite some time ago that downsampling would benefit from linear gamma space filtering, and upsampling would perceptually benefit from doing that in non-linear gamma space...

I am not so sure anymore that enlargement (upsampling) generally benefits from being performed, say, through Lab (=L*a*b*) or other perceptual color space (with gamma different than 1), or through a sigmoidal color space (which is a bit too blunt a tool, at least in the published versions, to be used blindly even though it sometimes gives very good results).
If you are using an upsampling method that does not create strong halos, enlarging through linear light appears to be generally better at preventing color artifacts.
With an upsampling method that adds overshoots and undershoots (like Lanczos filtering), it appears that using a perceptual color space (or sigmoidal one) may often be better, like it may be when sharpening.
The advice given in http://www.imagemagick.org/Usage/filter/nicolas is not "wrong" but it is slowly going stale. Let's call it a starting point.
For example, there are two filters that I now quite like that are not on the list: EWA RobidouxSoft http://www.imagemagick.org/discourse-server/viewtopic.php?f=22&t=19823&p=109820#p109820 or actually any Keys cubic with alpha (C in ImageMagick) larger than the RobidouxSoft value but strictly less than the Robidoux value, and EWA LanczosSharpest 4 http://www.imagemagick.org/discourse-server/viewtopic.php?f=2&t=25268#p109493.

(Windows users: http://www.imagemagick.org/discourse-server/viewtopic.php?f=1&t=23850&p=101409#p101409.)

I wish I had the time for "round 2". But I don't: It requires a lot experimentation, comparison and discussion to put such things together. At this point I can only ask questions.

P.S. In other words, I wish my "extensive trials" had been more extensive.

[NicolasRobidoux]

Sometimes, however, just asking the question is a step forward: http://www.imagemagick.org/discourse-server/viewtopic.php?f=22&t=21804

[NicolasRobidoux]

...
1) although it's a weak effect, people comparing resizing algorithms should probably take pains to make sure they do all their resizings at the same gamma, and should report that gamma when they report the results.
...

Sections 2.6 and 6.6 of my former student Adam Turcotte's Masters thesis
THE EXQUIRES (EXTENSIBLE QUANTITATIVE IMAGE RESAMPLING) TEST SUITE: IMPACT OF THE DOWNSAMPLER, DIFFERENCE METRIC, TEST IMAGE, RESAMPLING RATIO AND COLOUR SPACE ON UPSAMPLER RANK
http://web.cs.laurentian.ca/nrobidoux/misc/AdamTurcotteMastersThesis.pdf

[NicolasRobidoux]

Hi Jeff,

I'm afraid that's not universally true, but I assume Eric knows that but also has to deal with complexity trade-offs of the current sharpening dialog (it could use a serious overhaul). Deconvolution makes more sense in a linear gamma space, edge enhancement makes more sense in a gamma pre-compensated space (to reduce differences in acutance on the light side and the dark side of the edges/lines), and it would benefit from a possibility to target lighter, medium, and darker tones because their contrast is different in non-linear gamma space.

He probably, amongst others, paid attention to the discussions on the ImageMagick board and website where it was concluded quite some time ago that downsampling would benefit from linear gamma space filtering, and upsampling would perceptually benefit from doing that in non-linear gamma space.

Professor Nicolas Robidoux also did extensive trials of various methods used by ImageMagick, including temporarily (sigmoidal) contrast adjusted upsampling, especially on the more advanced Elliptical Weighted Averaging (EWA) filtered approaches but also on the tensor based variants of Lanczos windowing.

Cheers,
Bart

I wish I could take credit for inspiring Lightroom improvements but as far as purposely using something else than gamma 1 when upsampling I started openly pushing the idea in the Spring of 2012 (quite emphatically in http://www.imagemagick.org/discourse-server/viewtopic.php?f=1&t=20992) which is after the release of Lightroom 4.

[Jeremy Roussak]

Aim low, and you'll shoot something low, aim higher and you'll shoot something higher ... (metaphorically speaking, I do not endorse the use of guns )

I assumed you were referring to cameras 

Jeremy

[NicolasRobidoux]

...
Deconvolution makes more sense in a linear gamma space, edge enhancement makes more sense in a gamma pre-compensated space (to reduce differences in acutance on the light side and the dark side of the edges/lines), and it would benefit from a possibility to target lighter, medium, and darker tones because their contrast is different in non-linear gamma space.
...

P.S. http://www.imagemagick.org/discourse-server/viewtopic.php?f=1&t=25736