Pages: [1] 2   Go Down

Author Topic: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate  (Read 5914 times)

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267

Here's an image that may help people that analyze crosstalk effects, and why crosstalk is so subject dependent.

It's showing a 40HR IQ3100 18mm up landscape orientation, and the upper edge of that. This is bad stuff, green1 and green2 differs 38%. The crops are tiny section blowed up 800%, where talking pixel peep here! Then I show the raw image before demosaicing, that is we see the sample values in the bayer array.

You can see in the LCC shot that greens above blue pixels are much darker than the greens above red pixels. Why? Because we look were we look there's crosstalk flowing down and up in the image, that is reds crosstalks into greens above, and as the light has more red than blue, the greens above reds gets a stronger signal than the greens above blue. If there would have been no crosstalk both green1 and green2 would be the same in the LCC shot. I guess you can see this in rawdigger, or otherwise in RawTherapee (just disable demosaicing). (Note that there's of course also crosstalk from greens to reds and greens to blues too, all channels mixing up meaning desaturation and reduced tonality.)

Then we see the detail from the actual shot. It has about the same difference betwene green1 and green2. Why? Because the detail is white, about the same as the LCC card so the color coming in at this detail is close to what's coming through the LCC card.

This means that when we apply the LCC shot most of the green1/green2 difference is cancelled out, and as we see the demosaicing succeeds quite okay, the small brighter/darker pixels is probably false detail from the residual green1/green2 difference.

This shows an example where the image survives pretty well despite huge amounts of crosstalk, and I hope I've succeeded describing why. If instead of a white detail we had a saturated color the LCC shot would not succeed in evening out the green1/green2 difference and we would get mazing.

Unfortunately none of the test shots I got in this series actually got any colors in them so I couldn't show the failure case.

And I think we have a problem here. I'm not sure many have realized how this issue works, and that it is subject dependent. This means that you could run a series of tests with mostly neutral colors in the "red zones" and voilà, it looks quite fine. And then when you later on shoot colors with some saturation you get failures...

If you just want to make a quick and easy test to see how bad it can get, first shoot an LCC shot, then shoot another LCC shot with a red (or blue) filter gel taped to it (so you get a red uniform image), and then you apply the LCC shot to your plain red image. If you don't have any crosstalk the color cast will be cancelled out just fine, otherwise it will fail. What this test shows is what happens with a pure red object in any position of a real image.
Logged

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #1 on: January 11, 2016, 02:28:34 pm »

There are demosaicers that merge green1 and green2 so they are immune to mazing, VNG4 is one such, but it costs detail and you can get artifacts around small details, and you still have the desaturation issue of course.

It is possible to measure green separation and merge the greens if separation becomes too large (essentially run VNG4 in those areas), I do that in my own Lumariver HDR LCC algorithm. In theory you could also use the green separation as an input parameter to increasing saturation back up, but it would be very hard to make it match up perfectly in an even sky for example, so it's probably better to let it desaturate.

I don't know what type of tricks, if any, Capture One does to hide the crosstalk issue.

Again the red-filtered LCC shot is very effective in testing what happens when a saturated color comes into play.
Logged

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #2 on: January 11, 2016, 02:42:43 pm »

As mentioned I don't have any test files for IQ3100 to show this, but I dug up an old extreme test of a 44x33 crop 40 megapixel dalsa on extreme shift on SK28XL. Unlike the IQ3100 this Dalsa is worse in portrait orientation (so it's a worst case test!).

Here's a red-filtered LCC shot which has been color cast corrected with the corresponding real LCC shot. As you see there are quite some residual cast left. When levels go up huge there's also color shift, not only desaturation. The red filter used here is pretty strong cutoff so it's worse than real colors, one should probably use a weaker filter to make more real color result. If one is interested one can test a few different gels and see which colors that are hurt more than others.

The problem with any random real scene shot is that it very often happens to be close to neutral colored detail-free surfaces in the critical areas. Do you want to test the best case, or the worst case? Or the "common" case, and hope you won't come across the worst case? You can do as you like, it's your money ;). The filtered LCC shot is anyhow an easy and effective way for those that wants to test the worst case.

Note that sensors perform differently depending on orientation, so do your tests with the sensor in both portrait and landscape mode, ideally all four orientations. And to make things worse/messier the offset microlenses of the Sony sensors means that you should step say 5mm at a time as you may get a sharp performance falloff at some point which you may miss if you make huge steps.
« Last Edit: January 11, 2016, 02:52:14 pm by torger »
Logged

voidshatter

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 400
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #3 on: January 11, 2016, 03:03:00 pm »

Thanks Anders! What about the 15mm shift and 10mm shift? I think mazing artifact starts from 15mm and beyond but 10mm should be pretty safe.
Logged

Wayne Fox

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 4237
    • waynefox.com
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #4 on: January 11, 2016, 04:25:17 pm »

Is this issue you are discussing the reason for the current "bug" with LCC correction and green shift, and what 9.0.3 is supposed to resolve?  Is it bad enough you feel software changes won't "fix" it, or at least won't fix it some of the time?
Logged

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #5 on: January 12, 2016, 02:03:57 am »

Thanks Anders! What about the 15mm shift and 10mm shift? I think mazing artifact starts from 15mm and beyond but 10mm should be pretty safe.

I shall take a look later, hopefully tonight!
Logged

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #6 on: January 12, 2016, 02:34:30 am »

Is this issue you are discussing the reason for the current "bug" with LCC correction and green shift, and what 9.0.3 is supposed to resolve?  Is it bad enough you feel software changes won't "fix" it, or at least won't fix it some of the time?

It's a problem that's very hard to fix in software. I've myself tried to make a crosstalk cancellation algorithm when the IQ250 was new and this problem became more evident than ever, so that's why it's a "pet issue" for me.

What I did back than was to try to mathematically solve the issue, that is shift the crosstalk backwards. The problem is that you need to start from an area without crosstalk (ie center of the lens) and then calculate backwards from there, and when you chain several thousands of pixels like that the small errors and it becomes impossible to get a stable algorithm.

To "fix" it in software you instead need to hide it. It's easy to look into the LCC shot and see where green1 and green2 separates, and then equalize greens in those areas (at the cost of detail), then mazing will never occur. I do that already in my own LCC algorithm, but I don't think Capture One does it yet as I've seen reports of mazing still. If you do that you only have the color and tonality issues left, that is as channels are mixed up saturation and color separation goes down and there can be some hue shift.

It's a matter of taste how large issue you think this is. If you want to you can see it as a form of vignetting, "tonality vignetting" (it's not the same circular shape though!), and even think of it as a subtle creative effect. Or you can see it like this -- "why am I spending $44k on a back which is not designed for these lenses, and even think about compromising tonality?".

I think both conclusions are valid, I just want people to know what they are doing. I do find it surprising that many tech wide users are so obsessed by sharpness, and don't seem to worry much at all about tonality. What would you rather have, perfect color/tonality or perfect sharpness? We make hugely technical analysis of sharpness and DR, and I think this issue also deserves attention. It does get it but I think the issue of subject-dependency is a bit lacking. By explaining it in more detail in this thread I hope other technical-minded analysts that actually have access to this back (I don't) can make their own more detailed investigations.

Standard color cast is a calculated issue which can be fully neutralized with the simplest of LCC algorithms, it's as simple as neutralizing vignetting. This was an intentional tradeoff to allow for symmetrical and weak retrofocus lenses which is the foundation of the unique tech wide angle performance. Rodenstock designers did not have the intention to push the sensors into large amounts of crosstalk though, but they couldn't foresee what sensors would appear in the future. If they had known about where sensors would go the wide angles would not have been designed the way they are today. There is a critical issue here though, the more the tech wides are forced into retrofocus the harder it becomes to retain their unique performance. At some point it becomes unfeasible.
Logged

voidshatter

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 400
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #7 on: January 12, 2016, 07:39:05 am »

Given that the 23HR and the 40HR were announced about 8 years ago, I think it's already impressive that they could survive the 100MP Sony CMOS.
Logged

BernardLanguillier

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 13983
    • http://www.flickr.com/photos/bernardlanguillier/sets/
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #8 on: January 12, 2016, 07:51:49 am »

Given that the 23HR and the 40HR were announced about 8 years ago, I think it's already impressive that they could survive the 100MP Sony CMOS.

I would be great if Rodenstock announced a new generation of lenses at the Kina wouldn't it? ;)

No insider unformation here, just wishful thinking.

It seems overall that the 23 and 40 are pretty OK even without Center filter, I wonder how good the 32mm is with center filter and C1 Pro 9.03. I guess that Doug will let us know as soon as 9.03 becomes available.

Cheers,
Bernard

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #9 on: January 12, 2016, 08:48:09 am »

8 years should be a quite short time for a lens, but perhaps we'll have faster cycles now, I don't know. The lenses can start work better again if next CMOS generation is some sort of BSI, but I'm afraid the continuing chase for even higher megapixel numbers will keep the angular response narrow even if pixel technology improves. Phase One / Sony has also shown once again that the tech wides are not prioritized, probably too small market, so I don't think we will see any active attempt to support them in the future either. If it happens it would be due to a side effect of an improvement made for a different purpose.

Center filter cannot reduce crosstalk issues of course, but reduces vignetting. I don't see the vignetting issue as particularly large considering the very large dynamic range there is of the camera. The center filters can be a bit of a mess to shoot with on ultra wides due to reflections, so if the sensor has good DR it may be better to shoot without.
Logged

Paul2660

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 4066
    • Photos of Arkansas
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #10 on: January 12, 2016, 09:02:40 am »

I can't imagine shooting the 23mm without a CF, as even on center the vignetting/noise with at least a CCD back is considerable.  True also with the 28mm.  Remember the exposure compensation on the CF for the 23mm and 28mm is 2.5 stops, quite a bit.

The 32mm does fine without the CF on center, but I feel on shifts it helps quite a bit.

Surprised to see the 32 having as much an issue with crosstalk on this sensor as it did very well on the 50MP chip as I recall.   

Paul C
Logged
Paul Caldwell
Little Rock, Arkansas U.S.
www.photosofarkansas.com

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #11 on: January 12, 2016, 09:18:04 am »

Yes take my words about the DR thing as a word from the "low DR guy"... I'm pretty pleased making pictures with my Kodak CCD, which by today's standards feels like the noisiest sensor ever made, and to me those extra stops from the Sony sensor provides a huge margin for my shooting style.

Here's a low DR image for all of you to enjoy (not a Rodenstock lens though) ;)
Logged

Ken R

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 849
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #12 on: January 12, 2016, 12:38:12 pm »

Yes take my words about the DR thing as a word from the "low DR guy"... I'm pretty pleased making pictures with my Kodak CCD, which by today's standards feels like the noisiest sensor ever made, and to me those extra stops from the Sony sensor provides a huge margin for my shooting style.

Here's a low DR image for all of you to enjoy (not a Rodenstock lens though) ;)

Exactly, if someone would just open up some of the discussion threads and read them out of context or without knowledge about photography and gear in general they would think that the gear in discussions is junk, useless. That Phase 1 backs are worthless and that anyone who buys them is an idiot. Only safe buy might be the A7RII or D810. Dare anyone speak badly about the sensor in those two.  8)
Logged

voidshatter

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 400
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #13 on: January 12, 2016, 01:10:05 pm »

Exactly, if someone would just open up some of the discussion threads and read them out of context or without knowledge about photography and gear in general they would think that the gear in discussions is junk, useless. That Phase 1 backs are worthless and that anyone who buys them is an idiot. Only safe buy might be the A7RII or D810. Dare anyone speak badly about the sensor in those two.  8)

I can speak badly about these two sensors (in the context of landscape, cityscape, astrophotography etc) ;D

A7R-II: poor dynamic range for long exposure: http://www.dpreview.com/forums/thread/3890417
D810: no access to the Canon 17mm TS-E
Logged

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #14 on: January 12, 2016, 01:37:51 pm »

About the 40hr, here's some heat maps. The pixel values are in 0-100 range (hence a bit dark) and if the value is 90 it means green1/green2 (or green2/green1) is 0.9 (10% difference) which is the measure I've used for crosstalk.

Green is zero issues (less than 1.5%)
Yellow is up to 10%
Red is 10 to 30%
Blue is more than 30%

Back with the IQ250 minor mazing issues was seen for saturated colors with only 6% of crosstalk, but that was Phocus. Certainly with a DCB demosaicer it starts mazing with just a couple of percent crosstalk. I have tested to little with Capture One to know how robust their demosaicer/LCC is.

For reference I've attached the heat map of the test shot Dalsa/SK28XL above, so you can see some connection between crosstalk and color error.

Anyway, assuming that the demosaicer can deal with it my own judgment is that less than 5% is perfectly safe unless it's a really color critical situation. I would use it myself.

At 10mm shift the 40hr reaches 11% in the hot zone, it's probably not going to cause any problems in 95% of your landscape images, but if I would shoot architecture with colorful details all over I would not consider it a good solution for professional production, but that's my personal opinion. I would rather shoot an IQ360 for sure on that than risking having to manually clean up trouble images from time to time.

I think 15 and 18mm shifts are way past the limit what anyone should accept, more than 30% crosstalk. Just look at the reference image how much color issues you can have.

(If anyone wonders why I got the 50 megapixel old-school Kodak - it has no more than 2% crosstalk for extreme shifts on the badass SK28XL. Why? It was designed to handle wide angular response. It has lightshields between pixels, and it has no microlenses.)
Logged

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #15 on: January 12, 2016, 01:48:00 pm »

If someone wonders about the particular pattern where it seems like crosstalk is reduced back to a thin green line and then gets worse again it's a limitation of the measurement. Only green1/green2 difference is checked and as crosstalk flows between all pixels you can have a temporary situation when the difference is evened out somewhat, so in reality actual crosstalk is probably increasing all the way. That is the heat map can underestimate (but not overestimate) actual crosstalk.

The heat map is very good for detecting where "mazing" can occur though, as green1/green2 difference is causing that effect.

When it comes to showing tonality and color shift issues it's better to make that color-filtered LCC shot, but I almost never have such shots as I have done test shots myself, and with only a plain LCC the heat maps are the least bad estimation of crosstalk I can do.
Logged

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #16 on: January 12, 2016, 01:51:45 pm »

I've heard that the A7R-II is actually a bit disappointing concerning tech wide compatibility despite that it's a BSI sensor, but I haven't seen any hard facts on that.
Logged

torger

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3267
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #17 on: January 12, 2016, 02:11:03 pm »

If you have a tough requirement of max 5% or even 10% "crosstalk" (actually green separation in my heat maps) it's not looking good for the wides. The 23HR reaches 14% unshifted, the 32HR 20% on 10mm shift, the SK60XL which is tougher than many think as it's symmetric pushes the IQ3100 to 11% with 10mm shift.

The IQ360 is below 5% for all those.
Logged

ErikKaffehr

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 11311
    • Echophoto
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #18 on: January 12, 2016, 02:24:56 pm »

Hi,

I may have heard something like that, too. Not seen any real discussion on that issue.

I have an A7rII but I have only retrofocus lenses. Anyway, I would guess a smaller sensor makes less demands as it probably will be used with less shifts.

Chris Barret seems be pretty happy with his, but my understanding is that he uses it with Canon TSEs and V-series Hasselblad lenses. Pretty much as I do.

Best regards
Erik

I've heard that the A7R-II is actually a bit disappointing concerning tech wide compatibility despite that it's a BSI sensor, but I haven't seen any hard facts on that.
Logged
Erik Kaffehr
 

Paul2660

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 4066
    • Photos of Arkansas
Re: IQ3100 and tech wides: why crosstalk effects are difficult to evaluate
« Reply #19 on: January 12, 2016, 02:55:32 pm »

I guess the big question is, will you see it in post?

The IQ150, when shifted to 15mm with a 40 HR-W was basically pure red on the last 6mm or so of shift so much that when I first viewed it, I was sure that the 50MP chip just would not work.  However when the LCC was applied it's not bad at all.  In fact, what I found, was that the 40 HR-W had more issues with my IQ260 on a blue sky than the 50MP IQ150.  C1 has to me a lot of trouble getting a pure blue to work on a shift from the 260 as there always seems to be a residual red cast, which has to be fixed. 

The 150 only really showed me one issue on shifts, the light/faint blotchy vertical banding on solids.  This also showed up on the IQ3100 on the Alpa test where they shot a blue sky. 

I guess to me, it's not a show stopper, wasn't on the IQ150, only the crop factor was really.  If you are working in Art reproduction I guess it might for sure.  But in my outdoor work, I was very pleased with what Phase One was able to accomplish with the IQ150/250 and movements.

I am of course a bit disappointed that since all this was known about from the 150/250, that Phase did not attempt a better fix for the wides, but I guess that is just not possible?   Maybe the next chip? 

But from what I have seen on of the LCC's from Capture Integration on the 10mm shift with the 32 Rodie and 15mm of shift on the 40, the LCC's outwardly do not show as much red color, no where near as much, when compared to the IQ150 LCC.  Right now it's also hard to really compare with the LCC's since C1 is not processing the LCC for the 3100 without the green cast.

I realize that the LCC doesn't correct crosstalk, but on the 50MP chip it did a darn good job of getting all the red out and bring back the correct looking colors.

Just one opinion of many.

Paul C

Logged
Paul Caldwell
Little Rock, Arkansas U.S.
www.photosofarkansas.com
Pages: [1] 2   Go Up