theSuede wrote:
Please don't forget that the LX-3 sensor is manufactured in equipment that DSLR-sensors can only DREAM about, and then add in that surface efficiency is a lot higher in those manufacturing lines. We just hit 180nm in the latest sensors (1D4, 7D, the latest D3s and panasonics latest) - that's the same level of manufacturing resolution as the small sensor business have had available for 6-7 years now. We just saw the first sensors incorporating Cu leads in stead of Al.
The 7D sensor manufactured in LX3-class specifications would quite probably make the 7D more expensive than the 1Ds3. Cost = area squared divided by manufacturing technology maturity.
Surface efficiency is very important, as silicon crystal has a very high refractive index - 3.6 something for average green light - compared to "high" refraction special glasses used in lenses - normally around 1.8. So, as long as you can get the light to hit the sensor surface at the right place, light is bent to pass almost straight down into the sensor, no matter what the angle of incidence was. You have to get the light there in the first place - and this is a lot easier in a small die sensor. They have a lot thinner filter elements, and some other manufacturing advantages too that add to the surface efficiency....Show more →
theSuede wrote:
The BEOL-view of the 1.75u sensor in the chipworks link is interesting - compare the scales of the cell with a larger cell sensor... They're not that different! the ratio of cell width to package layer height is almost the same - which explains most of why the small sensors can be made so light sensitive. It also explains why their colour discrimination capability is no better at ISO80 than some FF/FX-cameras can manage at ISO3200... :-) Are we willing to let our cameras loose 90% of their colour resolution capability for "high-resolution" purposes?
Could you explain in more detail this last statement? I would have said that digicams at ISO 80 are like DSLR's at ISO 1600, simply because over the frame both are collecting the same number of photons (ratio of sensor areas compensates flux ratio).
Since it's hard to make a filter material very dense optically without at the same time making it very uneven in physical quality (which is a no-no in semiconductor manufacturing), the (lack of) thickness of the CFA plate indicates at least a doubling of transmittivity. If the used materials are the same, then scale dictates that the filter plate thickness should be at least two times thicker than it is depicted. This would make the cell filterplate at least twice the depicted height, and hence VERY sensitive to angle errors. In short - the sensor would be totally useless with anything below 100mm (unless perfectly retrofocal) or F/4.
And rest assured, if a new, super-filter-material had surfaced, larger sensors would have received it's blessings at the same time as smaller sensors.
High filter transmittivity is good for noise characteristics, and VERY bad for colour accuracy. To use a very recent example, the 5D (original) can accurately discriminate between twice as many individual hues per frame as the 5D2, even though the 5D2 has a higher signal to noise ratio per area unit (at the same exposure) AND higher resolution. All due to filter dilution, to get lower noise. The greens of the 5D2 are twice as diluted as in the 5D, and the other two primaries got their share with the ugly-stick too. This leads to quite a few problems with metamerism, and an uneven hue transition curve in between the primaries (most commonly known as "Canon skin tones"). The colour profiles used with the 5D2 in raw converters needs to be quite a lot sharper and more radical in their corrections than the ones used for the 5D. And higher resolution corrections invariably leads to very much worse "worst case" errors when the profile is used on WB-temperature outside the (in this case) very narrow "best use" window. The latest X-rite "Passport" only corrects for six points around the HSL hue-circle, and this is not nearly enough to ply the colours of the 5D2 back into shape. 36 would be more appropriate.
Example in case:
I can make a reasonably accurate profile for a 5D in very difficult fluorescent lighting quite easily. This is almost impossible with the 5D2. The error margins will be higher than the correction accuracy.
Now lower the colour-accuracy of the 5D2 by a factor of three, and you have a working 2um cell.
theSuede wrote:
High filter transmittivity is good for noise characteristics, and VERY bad for colour accuracy...
...This leads to quite a few problems with metamerism, and an uneven hue transition curve in between the primaries (most commonly known as "Canon skin tones")...
Thank you for the very interesting discussion!
It makes me glad that I am "color blind." -- 'yeah, that skin looks fine to me...'
In addition to the 5D vs 5D2 example you give, what can be said about the various Canon dSLRs and their color quality (hopefully subjectively)? I've read many people praise the D60 (or was it the D30?; both are before my time, dSLR-wise) and the 1D - do these nuggets of perception square up with rigorous quantitative analysis?
I notice that it does not exhibit the changing noise performance that I noticed with my 1Ds3 if you select third stop ISO increments as opposed to whole stops.
Joakim (theSuede), your description of the situation regarding CFAs is the first cogent and sensible argument I've read so far on this board for how higher pixel density can be detrimental to image quality. Well done.
thedigitalbean wrote:
Joakim (theSuede), your description of the situation regarding CFAs is the first cogent and sensible argument I've read so far on this board for how higher pixel density can be detrimental to image quality. Well done.
I have been reading the Suede's post with great interest recently.
Thank you Joakim for bringing new thought and argument into the discussion.
thedigitalbean wrote:
Joakim (theSuede), your description of the situation regarding CFAs is the first cogent and sensible argument I've read so far on this board for how higher pixel density can be detrimental to image quality. Well done.
Sorry, I am not entirely convinced by Suede's postulate.
I just used DXOMark to compare the color sensitivity of Nikon D3X (24 MP FF) vs Canon 5D (12 MP FF) vs Canon 5D2 (21 MP FF). The performance ranking of the cameras goes as 5D > D3X > 5D2. But the difference is VERY SLIGHT.
Perhaps this has more to do with the RAW conversion program Suede uses?
On a slightly different but related note, Jeff Ascough has this to say about skin tones:
"Nikon approached me in June 2008 about using their gear. I took a D3 on loan shot two weddings with it, and sent it back; I didn't like the skin tones. At high iso they looked artificial and digital... As I understand it Canon use a different method of onboard NR to Nikon, which gives Nikon the edge in terms of actual noise but at the expense of skin tone at high iso."
This assumes we believe Jeff who also says: "Think about it for a minute, if I think a camera is crap but I say it is fantastic, then don't you think everyone will find out once they buy the camera? Then my reputation is down the toilet."
theSuede wrote:
Since it's hard to make a filter material very dense optically without at the same time making it very uneven in physical quality (which is a no-no in semiconductor manufacturing), the (lack of) thickness of the CFA plate indicates at least a doubling of transmittivity. If the used materials are the same, then scale dictates that the filter plate thickness should be at least two times thicker than it is depicted. This would make the cell filterplate at least twice the depicted height, and hence VERY sensitive to angle errors. In short - the sensor would be totally useless with anything below 100mm (unless perfectly retrofocal) or F/4.
And rest assured, if a new, super-filter-material had surfaced, larger sensors would have received it's blessings at the same time as smaller sensors.
High filter transmittivity is good for noise characteristics, and VERY bad for colour accuracy. To use a very recent example, the 5D (original) can accurately discriminate between twice as many individual hues per frame as the 5D2, even though the 5D2 has a higher signal to noise ratio per area unit (at the same exposure) AND higher resolution. All due to filter dilution, to get lower noise. The greens of the 5D2 are twice as diluted as in the 5D, and the other two primaries got their share with the ugly-stick too. This leads to quite a few problems with metamerism, and an uneven hue transition curve in between the primaries (most commonly known as "Canon skin tones"). The colour profiles used with the 5D2 in raw converters needs to be quite a lot sharper and more radical in their corrections than the ones used for the 5D. And higher resolution corrections invariably leads to very much worse "worst case" errors when the profile is used on WB-temperature outside the (in this case) very narrow "best use" window. The latest X-rite "Passport" only corrects for six points around the HSL hue-circle, and this is not nearly enough to ply the colours of the 5D2 back into shape. 36 would be more appropriate.
Example in case:
I can make a reasonably accurate profile for a 5D in very difficult fluorescent lighting quite easily. This is almost impossible with the 5D2. The error margins will be higher than the correction accuracy.
Now lower the colour-accuracy of the 5D2 by a factor of three, and you have a working 2um cell....Show more →
I think this is very interesting, although I'm not sure how much reality it reflects. I do indeed notice a certain fragility in 5D2 colour, and often a sense that the colour palette is somewhat limited, particularly in the deep shadows where a green/magenta duality seems to take over.
However, I do get great results on skin from my 5D2's by applying my own profiles, which are suited to different kinds of picture, and I don't find that the 5D competes generally in overall image quality, including the colour characteristics.
My 1Ds3 colours felt a little more robust in processing, which illustrates that this is not a problem inherent to high density sensors, but might instead be only a question of the colour filters. My 1Ds3 was great at iso 12800.
Limited colour palettes were what gave films like Kodachrome their beloved warmth. If you walk around the Louvre you won't find a priceless painting which contains as wide a gamut as the 5D2 can capture. The 5D2's colour characteristics must be wielded judiciously, but they can be. I certainly hope that future cameras have a more accurate and uniform response to different colours of light source, but for the moment we must make the best of it.
I like Jeff Ascough, he's a good photographer, but really..... he didn't like the D3... hahahahaha.
That's like Joe McNally saying he prefers Nikon. Shock horror!
Quote from Canons site:
Canon Ambassador Jeff Ascough is regularly voted as one of the world’s top wedding photographers by photo experts on both sides of the Atlantic Ocean. Although he has been an habitual Canon EOS-1Ds Mark III user, for his regular wedding work, CPN gave him a new EOS 5D Mark II to see how it suited his field of photography.
Jeff never liked the Canon 5D because he thought it was flimsy, compared to his 1 series. But give him a 5DII and he's a convert.
thedigitalbean wrote:
Joakim (theSuede), your description of the situation regarding CFAs is the first cogent and sensible argument I've read so far on this board for how higher pixel density can be detrimental to image quality. Well done.
I'm not sure that Joakim has made the argument that you think he has made - that is, I think you are "connecting the dots" that he laid down, but it may not be correct to connect those dots.
theSuede talked about how "High filter transmittivity is good for noise characteristics, and VERY bad for colour accuracy." A few posts earlier, he talked about how quantum efficiency, and "The angle problem is starting to show up in both the 50D and the 7D..."
It seems that an unanswered question is: was Canon "forced" to use a more-transmissive, lower-color-quality CFA for the 7D and other high pixel density sensor cameras to compensate for Qe losses attributable to high pixel density in order to maintain net system light-recording efficiency? Or do the new filters improve sensitivity more than any loss due to pixel density increases?
Reading the technical discussions that Canon posted about the 7D sensor (to the extent that it is accurate), it seems that there are a number of changes/"improvements" made relating to efficiency. One that seems relevant to "the angle problem" mentioned by theSuede is that the distance between microlenses and the photodiodes was reduced. Hmmm...
I'm not saying that I agree with theSuede or buy his argument completely. I was just saying that the way it was presented (with some empirical observation), much like brainiac's argument is presented was something I appreciated.
thedigitalbean wrote:
I'm not saying that I agree with theSuede or buy his argument completely. I was just saying that the way it was presented (with some empirical observation), much like brainiac's argument is presented was something I appreciated.
Thanks for clarifying, and sorry if my tone was unduly critical. I too appreciate argument backed with evidence and careful reasoning.
next time just skim. 
