p.2 #1 · High-Iso Performance vs good colour rendition
You know, doing a little research at DxO as to what this number is really about leads me to believe we may be putting far more importance on it than we should. DxO themselves state this about the measurement:
"In practice, the SMI for DSLRs ranges between 75 and 85, and is not very discriminating. It is different for low-end cameras (such as camera phones), which typically have a SMI of about 40. For this reason, we give this measurement as an indication but do not integrate it in DxO Mark."
The more important number with regard to color would appear to be the color sensitivity number represented by the Portrait Score, not the color response number we have been discussing.
"The best image quality metric that correlates with color depth is color sensitivity. Color sensitivity indicates to what degree of subtlety color nuances can be distinguished from one another, often meaning a hit or a miss on a pantone palette. Maximum color sensitivity reports, in bits, the number of colors that the sensor is able to distinguish.
The higher the color sensitivity, the more color nuances that can be distinguished. As with dynamic range, color sensitivity is greatest when ISO speed is minimal, and falls rapidly with rising ISO settings. DxO Labs has focused on measuring only maximum color sensitivity.
A color sensitivity of 22bits is excellent, and differences below 1 bit are barely noticeable."
p.2 #2 · High-Iso Performance vs good colour rendition
That the Nikon cameras (the D800 and the D7000 is indeed almost identical, they differ less in color than my lowest possible measurement error...) get a lower score is indeed a conscious choice by Nikon. They chose the CFA construction, to their requirements. Their exact reasoning for this is beyond me, to get an answer you'd need to speak directly with their design group, and I doubt they would give you a straight answer... :-)
The biggest Nikon disadvantage is just like the medium format and Leica disadvantage, they actually have to much color separation in some hues, which can look good in some lights and worse in other lights. What separates Nikon from Leica / MF is that their filters are almost perfectly aligned to minimize color noise at higher ISOs.
My guess is that this is what they were aiming for - more color separation on the sensor means less color amplification in the raw conversion, and that means less color noise in the developed image.
The DxO MSI index is based on an ISO standard, and DxO use their own interpretation of a variation of this. They use the Gretag colorchecker (classic 24p version) and the raw ADU values from the camera. Basically this means that they use a "channel mixer" just like the one you have in photoshop - directly on raw values as input - and then they run an optimization script on the settings in the channel mixer to get the mixer setting that has the lowest average dE value compared to the measured reference.
The DxO "result" is then calculated as: 100 - (average dE * 5.5).
I have no idea where the constant "5.5" comes from, it's probably just an arbitrary constant meant to give meaningful results in the range from 0dE to 10dE. 0dE = score of 100 and 10dE = score of 45.
A result of 80 means that the average dE was (100-80)/5.5 = 3.6 dE
This is not much different from say a value of 85; (100-85)/5.5 = 2.7 dE
Their metric isn't defined, but if they follow the ISO standard the dE value should be in CIE1976 Lab, with no psychovisual adaptations. This means that a score of 82 can actually be "visually" worse than a score of 81, depending on how the errors were distributed...
A camera profile does however almost always contain a LUT part too, and with a target as "easy" as a CC24 you can easily get a score of "100", perfect. This perfect score will however fall apart when you include more colors, it's only valid for the limited set of colors that the reference chart contains. Actually it's not about the amount of colors on the chart, it's all about the amount of individual pigments used to mix the colors on the chart.
If you print a colorchart with 1000 variation patches in hue, sat and lightness in a normal inkjet printer, it will still only contain four or six individual pigments (the amount of individual pigments present in the inks available in the printer). This means that the color chart you printed is only usable for checking the printer in itself - it's basically useless to take a photograph of it try and calibrate a camera with it - this will tell you very close to nothing. To few variations, and much to sensitive to light metamerism (how the light you use interact with the printer inks).
p.2 #3 · High-Iso Performance vs good colour rendition
^^^ I've read the DxO write-ups, and do not really understand how color sensitivity increases with downsampling an image. I understand how signal-to-noise and thus dynamic range (ability to extract information from the shadows) improves. But how does improved color sensitivity come about? In fact, it increases faster with downsampling than dynamic range.
p.2 #4 · High-Iso Performance vs good colour rendition
Yeah, that's why I think we need theSuede to weigh in. If we're going by Color Depth scores, both the A900 and 5dii get a 23.7, and the NEX-7 gets a 24.1. With the 5dii's cfa change from the 1dsiii, I'm not sure what to think.
p.2 #5 · High-Iso Performance vs good colour rendition
Very interesting. I would assume the poor Leica M9 score is down to the poor soft/firmware in-camera. Profiling for different types of artificial light to avoid really pink lips and mottley skin is essential.
EDIT: or the much more likely explanation theSuede gave
p.2 #6 · High-Iso Performance vs good colour rendition
Doesn't DxO calculate some direct color separation numbers? If you pick a camera and go to Measurements -> Color Response, they give you RAW channel readouts for sRGB primaries. As is usual with DxO, their overall index is of doubtful usefullness, but you can look at the per-channel charts and see how color separation works for a given CFA.
p.2 #7 · High-Iso Performance vs good colour rendition
Arun Gupta wrote:
^^^ I've read the DxO write-ups, and do not really understand how color sensitivity increases with downsampling an image. I understand how signal-to-noise and thus dynamic range (ability to extract information from the shadows) improves. But how does improved color sensitivity come about? In fact, it increases faster with downsampling than dynamic range.
Maybe the word "sensitivity" was ill chosen, but I think it says what they meant; it's just not very clear what they meant... :-)
Basically the color sensitivity is the number of unique colors a camera can distinguish between when noise has been accounted for - in total, within the colorspace defined by the camera raw color.
If you take one defined mathematical point in a 3D color space, that point represents a "unique color". Now, depending on:
a) the color accuracy of the CFA, and
b) the amount of noise per channel in the raw
-each of those points will have a defined "error margin". A certain amount of noise 'N in one of the channels will mean that the point 'p can in reality be anywhere between 'p - (0.5'N) and 'p + (0.5'N). This goes for all three channels, so you get an "error volume", a small ellipsoid around the point 'p where all included colors are statistically valid options. The 'p can be anything within this smaller volume. Think of it as a very small "bubble" in the colorspace.
If you divide the total colorspace volume with the average volume of those "color error margins", you get the number of unique points where no point is within another point's error margin. Convert this to log2, and you get "bits" of color.
As it is a volume, you also get three dimensional noise errors - if you lower noise by 1Ev (same as a halving, 50%, or 0.5 in linear terms), each error margin ellipsoid volume shrinks to 0.5*0.5*0.5 = 0.125, or -3Ev. It shrinks in all three dimensions.
This is why the downsampling is so (theoretically) effective, and the DxO "color depth" grows faster with downsampling than the other metrics they use..
In reality though, pretty much nothing happens until you have scaled the original image down to 30-40% or less, since the spatial color resolution in a Bayer sensor is pretty low. It is very hard to accurately predict the color of ONE pixel, every pixel needs a surround support to get "color" in a Bayer based sensor. You can never get "pixel level" accurate color spatial resolution in a Bayer sensor. If you could, moire would not exist.
p.2 #8 · High-Iso Performance vs good colour rendition
'If we're going by Color Depth scores, both the A900 and 5dii get a 23.7, and the NEX-7 gets a 24.1'
and
'In practice, the SMI for DSLRs ranges between 75 and 85, and is not very discriminating.'
OK. Very reassuring. Nothing to see here, they are all the same WRT colour rendition, time to move on.
Seriously, thanks for the research and input, Tariq and Douglas, and of course theSuede. I stand with some other people who posted in this thread, like hauxon's friend and others, since it all seems to come down to using your own eyes. Nice to establish that much, anyway.
I have no idea why DxO even bother reporting data that means something between nothing and very little, but that is camera review sites for you.
Now, lets get back to something much more important, like noise at 25,600.
p.2 #9 · High-Iso Performance vs good colour rendition
@Hauxon, thanks for sharing that info from your friend. That backs up what I was seeing too with the D800 from looking at the limited number of shots recently posted on the net.
Why can't Canon/Nikon make two identical cameras like say D800 but have one with a thick CFA for low iso/great color optimized for landscape/studio shooting and then another with thin CFA for high iso shooting with not as good color? Would it not just be a different CFA/sensor assembly? How are the CFA's added to the sensor?
p.2 #12 · High-Iso Performance vs good colour rendition
wayne seltzer wrote:
@Hauxon, thanks for sharing that info from your friend. That backs up what I was seeing too with the D800 from looking at the limited number of shots recently posted on the net.
Why can't Canon/Nikon make two identical cameras like say D800 but have one with a thick CFA for low iso/great color optimized for landscape/studio shooting and then another with thin CFA for high iso shooting with not as good color? Would it not just be a different CFA/sensor assembly? How are the CFA's added to the sensor?
That would be lovely. CFA is either cemented to sensor, or simply held there same way as AA and IR filter. MaxMax did some surgery on 5D I think and converted it to BW camera by removing CFA. Tho I dont think it would be exactly easy to fit new CFA on it.
But manufacturers shouldnt have this problem. Thing is, that I dont think they care much about color performance. Cause most ppl are simply ok "if colors are good" and most important right now is megapixel and ISO war. I dont think we will see color war.
Bits vs metamerism index in reality..
Well, Sony A900 has higher metamerism index than D3X, right? And 1 bit less in color depth than D3X. How does it work in real world, anyone saw better colors from D3X?
Sure, bits show how much color info you can get from camera and how smooth will be gradients, but how usefull is having more bits of color information, when colors are "not right"? Its bit like ISO 102400, lovely number, but limited use.
p.2 #13 · High-Iso Performance vs good colour rendition
Mescalamba wrote:
Sure, bits show how much color info you can get from camera and how smooth will be gradients, but how usefull is having more bits of color information, when colors are "not right"? Its bit like ISO 102400, lovely number, but limited use.
But apparently the ten point or so differences in the cameras we are discussing is all but insignificant according to DxO and Joakim (If I understood what he said correctly) with regard to the color response/ metarmerism index (at least how it's measured here). Sounds like these numbers don't really mean much for all practical uses/ purposes, unlike the portrait score (bit numbers) which do.
p.2 #14 · High-Iso Performance vs good colour rendition
FWIW, I know that Iliah Borg, who owns both D3x and A900, prefers the A900 for landscape, because of the better green separation, according to him. I reference him, because he does the color transforms for RPP, and is pretty knowledgeable in the field.
p.2 #15 · High-Iso Performance vs good colour rendition
douglasf13 wrote:
FWIW, I know that Iliah Borg, who owns both D3x and A900, prefers the A900 for landscape, because of the better green separation, according to him. I reference him, because he does the color transforms for RPP, and is pretty knowledgeable in the field.
p.2 #16 · High-Iso Performance vs good colour rendition
Tariq Gibran wrote:
But apparently the ten point or so differences in the cameras we are discussing is all but insignificant according to DxO and Joakim (If I understood what he said correctly) with regard to the color response/ metarmerism index (at least how it's measured here). Sounds like these numbers don't really mean much for all practical uses/ purposes, unlike the portrait score (bit numbers) which do.
Correct: They don't. The "bit depth" tells you about how noise interacts with color discrimination, the "metamerism index" tells you an average eye-camera color registration difference.
A higher RMI is undoubtedly better, but in what way is hard to say. It's just like the DxO "scores" - it's a fixed number that tells you much to little about what's really happening.
The advantage the A900 has in green is actually that it has a smoother red... :-) since hue placement is determined by how the color channels in raw "cross" each other's boundaries, smooth and regular slopes on both sides of the peak are a must to get good color gradation. Steeper slopes do give great color hue resolution, but within narrow color bands. Make the slopes to steep, and you get "blind spots", and Nikon typically have a "blind spot" between 545nm and 560nm, which is mid green. This is typically of no importance whatsoever to actual greenery, since neither chlorofyll A or chlorofyll B contain any "unique properties" in this band of wavelengths.
In fall, though, the yellows - yellow/greens lose a bit of hue resolution. Not something I would consider a large drawback.
Canon is typically the absolutely worst for green/yellow separation - slightly better than most compact cameras. Their red and green (in Canon's case actually more like "orange" and "warm green") overlap so much that it's hard to get a good separation.
I typically work with quite a lot of shadow detail, so neither the A900 or the A850 worked better than the D3x for me.
That link deserves its own thread!