Home · Register · Search · View Winners · Software · Hosting · Software · Join Upload & Sell

Moderated by: Fred Miranda
Username   Password

  New fredmiranda.com Mobile Site
  New Feature: SMS Notification alert
  New Feature: Buy & Sell Watchlist
  

FM Forums | Alternative Gear & Lenses | Join Upload & Sell

1      
2
       3       end
  

Archive 2013 · Bayer masks and DxO color depth measurements
  
 
philip_pj
Offline
• • • •
Upload & Sell: Off
p.2 #1 · p.2 #1 · Bayer masks and DxO color depth measurements


'I know people like colors from Alpha and Leica, but I am looking for something solid that shows that there is really better separation.'

Some things resist the best and worst efforts of 'scientists'. Minolta had a reputation for pleasing colours long before Sony bit the bullet and took a dfferent CFA path to the other big companies.

Sometimes the most solid thing is to trust your own instincts and in this case, eyesight.

We all see differently, there are plenty of anthropological studies providing evidence of this, as it even happens at the cultural level. Even one's native language affects perceptions of colour separation.

Finally 'Alpha' is not a good overall description any more, the a99 is quite different to say, the a900. I see it every day, in every image. Much more subtlety, tonal gradation, noise handling, highlight control, DR, 14 bit=no banding. People forget colour is very dependent on a host of other factors not typically measured in a narrow, perhaps scientific, assessment of the subject.



Mar 13, 2013 at 09:08 PM
Mescalamba
Offline
• • • •
Upload & Sell: Off
p.2 #2 · p.2 #2 · Bayer masks and DxO color depth measurements


M9 is aimed like all digital Leicas at pleasing colors, which for me works. But truth is that some ppl dont like them.

I think that medium format backs have advantage in having 16-bit ADCs, compared to 12 + something that most dSLRs can do. I wouldnt mind if someone tried to do regular dSLR with 16-bits.



Mar 13, 2013 at 09:09 PM
carstenw
Offline
• • • • • •
Upload & Sell: Off
p.2 #3 · p.2 #3 · Bayer masks and DxO color depth measurements


AFAIK many MFDBs also have pipelines which are effectively 14 bits, not 16.


Mar 13, 2013 at 10:18 PM
telyt
Offline
• • •
Upload & Sell: Off
p.2 #4 · p.2 #4 · Bayer masks and DxO color depth measurements


Mescalamba wrote:
I think that medium format backs have advantage in having 16-bit ADCs, compared to 12 + something that most dSLRs can do. I wouldnt mind if someone tried to do regular dSLR with 16-bits.


On another forum there has been considerable technical discussion of this topic and some claim that the base noise of sensors is too high to see any meaningful additional data in a 16-bit pipeline vs a 14-bit pipeline. I don't have the technical background to comment further on this subject.



Mar 13, 2013 at 10:46 PM
mpmendenhall
Offline
• • • •
Upload & Sell: Off
p.2 #5 · p.2 #5 · Bayer masks and DxO color depth measurements


telyt wrote:
On another forum there has been considerable technical discussion of this topic and some claim that the base noise of sensors is too high to see any meaningful additional data in a 16-bit pipeline vs a 14-bit pipeline. I don't have the technical background to comment further on this subject.


That's probably true. Here's a rough calculation estimate:

Suppose a sensor can detect between n=1 and n=N photons, with low enough readout noise that noise is limited by fundamental counting statistics of sqrt(n).

The smallest increment you'd ever need to measure on this sensor is 1 count; measuring any finer than the minimum noise of sqrt(1)=1, like in 0.3 count increments, would be silly and unhelpful. So, with a simple linear readout scheme, a 16-bit pipeline would cover a scene dynamic range from 1 to 2^16 counts --- a 16-stop DR scene (unusual to actually find in real life, and a bit beyond current easily available sensor technology); for photographing typical real-life scenes with <14 stops of dynamic range, there'd be no benefit to >14 bits readout even with a "perfect" sensor.

This linear readout scheme, however, is already a highly inefficient use of bits. While you need increments of 1 at n=1, for n=100 the noise is sqrt(n)=10, etc.: you can get away with representing a lot lower resolution steps at higher values, because the noise is higher. A maximally "efficient" use of pipeline bits only needs to represent integral(1/sqrt(n),n = 0 to N) = 2*sqrt(N) values. Thus, a 16-bit representation would cover scenes up to 2*sqrt(N)=2^16 => N=2^30: 30 stops of dynamic range! I'd recommend staying away from scenes with 30 stops of dynamic range, because being incinerated by a nuclear blast is no fun. For a 16-stop DR scene, you'd only need log_2(2*sqrt(2^16)) = 9 bits of "maximally efficiently used" processing pipeline. Of course, it's a lot easier to "waste" extra resolution by just processing 16-bit linear signals.



Mar 13, 2013 at 11:22 PM
carstenw
Offline
• • • • • •
Upload & Sell: Off
p.2 #6 · p.2 #6 · Bayer masks and DxO color depth measurements


IIRC the Leica M8 had some similar more efficient use of the bits in its raw format. It was 10-bit or maybe even 8-bit, but not linear. I don't recall the details, but do recall the very interesting discussions we had about it over on the l-camera-forum when it was discovered.


Mar 13, 2013 at 11:57 PM
theSuede
Offline
• • • •
Upload & Sell: Off
p.2 #7 · p.2 #7 · Bayer masks and DxO color depth measurements


There are a few different (but all interconnected) discussions going on here - at the same time.

But the original question, "Color Depth": This metric as per DxO definition is VERY convoluted unless you have quite a thorough mathematical understanding of both how a tristimuli color representation works, and how noise-covariance step integration works.
But basically it is the numerical count of [with statistical accuracy!] separable colors that the sensor can distinguish at the given ISO, given in log2 format (bits)

This is based on the uncertainty that noise gives. If a sensor has very low levels of noise, uncertainty and the noise covariance diameter decreases, so the next color that is "definitely not the same" lies closer. More noise means that you have to move away further in color before you can say that the color has actually changed.
If the camera's color filters are very weak (like in a compact camera or cellphone) then colors have to be boosted - and this boost amplifies noise by exactly as much as it amplifies saturation.

Imagine a small marble of color. The size (diameter) of the marble indicates how much noise that exact color is affected by. The noise (diameter) is photon shot noise + electronic noise, multiplied by the amount of color boost needed. The marble next to this one has to be at least one diameter away to fit, and it has its' own diameter as calculated for its' own center color. And so on.

Given that you have a marble bag (color space) that is "this big", how many marbles will fit into that bag? That is the color depth. Less noise, less color boost needed = smaller marbles = more marbles will fit into the same bag.

This has nothing (except by necessary cooperation - you can't have one without the other!) to do with color accuracy per se, it's just a measurement of how many colors the camera can distinguish from any other color within its' working gamut.

MDFB's often have very good (to good!) color separation, but often not that brilliant DR - and their over-the-top color separation needs tempering to conform to human vision. But when you combine very good color separation with very low mid-gray noise levels (something MDFB's are extremely good at, but isn't directly related to DR!), you get really impressive gradation possibilities. It may not be perfectly translated into human-vision color, but it does look impressive.
....................

Color accuracy on the other hand, is more about matching camera color to human vision color. Here DxO uses metameric indexes as in ISO standard 17321. This uses ONLY linear math to try and match camera raw's to human vision. It is a measurement of how well you can fit the CFA filters to the human eye response, without resorting to spot color corrections.

But spot color corrections are part of almost any camera more advanced than a cell-phone module nowadays, so the SMI should be taken with a big lump of salt. It doesn't tell that much about the camera, and it certainly does not tell if the errors produced are good-natured (easy to correct with spot color corrections) or pure metameric failures that are impossible to correct.

A camera with an SMI of 80 can be WORSE than a camera with an SMI of 75, if the errors are in sensitive colors and if they are incorrigible.
But a camera with SMI80 will almost certainly be a lot better than the camera with SMI60.



Mar 14, 2013 at 02:15 AM
snapsy
Offline
• • • • •
Upload & Sell: On
p.2 #8 · p.2 #8 · Bayer masks and DxO color depth measurements


theSuede wrote:
There are a few different (but all interconnected) discussions going on here - at the same time.

Thanks theSuede. Can you shed light on how DxO's "Full Color Sensitivity" tab and Lab map correlate to metamers and/or color sensitivity, and also how to interpret the Lab map? Thanks again.



Mar 14, 2013 at 03:30 AM
carstenw
Offline
• • • • • •
Upload & Sell: Off
p.2 #9 · p.2 #9 · Bayer masks and DxO color depth measurements


theSuede wrote:
Color accuracy on the other hand, is more about matching camera color to human vision color. Here DxO uses metameric indexes as in ISO standard 17321. This uses ONLY linear math to try and match camera raw's to human vision. It is a measurement of how well you can fit the CFA filters to the human eye response, without resorting to spot color corrections.

But spot color corrections are part of almost any camera more advanced than a cell-phone module nowadays, so the SMI should be taken with a big lump of salt. It doesn't tell that much about the camera,
...Show more

This is quite interesting to me. So what measurement can be made which takes these spot colour corrections into account, and thus gives an accurate real-world idea of how accurate the colours of a given camera are? Does DxO or some other site list such numbers?



Mar 14, 2013 at 07:35 AM
theSuede
Offline
• • • •
Upload & Sell: Off
p.2 #10 · p.2 #10 · Bayer masks and DxO color depth measurements


snapsy wrote:
Thanks theSuede. Can you shed light on how DxO's "Full Color Sensitivity" tab and Lab map correlate to metamers and/or color sensitivity, and also how to interpret the Lab map? Thanks again.

The Lab map in f.ex DxO: Sony NEX-6 shows you the combination of:
ISO setting, the noise the sensor gives at that ISO and indicated image brightness, translated into a full color image.
The noise and color inaccuracy effect is calculated for ONE pixel, and that pixel is situated in the middle of a flat surface of the target color indicated (no luminance detail, just a perfectly flat color surface). The ellipse shows the standard deviation of any pixel on that [flat] surface. The colors are the original Gretag Macbeth CC24 formulation.

The noise present combined with the CFA qualities gives that the (de-bayered and color-corrected) color in the pixel you're looking at will vary within the ellipse indicated. Bigger area ellipse = more noise, or in this case "color uncertainty". Noise reduction and downsampling will of course make the ellipse smaller again, but can have other unwanted effects - especially on detail.
It's also quite educating to see how a CIE "A" light (normal old type light-bulb, modern low-effect halogen) changes the noise covariance... Since there's very little "blue" light energy content in an incandescent bulb, blue has to be amplified by white-balance. Amplification = more noise. More noise in blue = the ellipses are elongated in the Lab "b" axis (blue>yellow).

To get a very small noise covariance, you need low noise and good color separation on the CFA.
This measurement is only valid on a global (color) scale! It's very hard, actually impossible, to see if there's something wrong in specific hues with this model. It shows the overall globally averaged color variance, as calculated by a THEORETICAL setup given by the "best case" color transform needed to get from raw to some standard RGB (actually standardized XYZ, since we're looking at a Lab data amp).

High SMI index cameras will have small practical vs theoretical differences in this metric. Low SMI index cameras can (will?) have large difference between theory and practice in this metric.



Mar 14, 2013 at 02:29 PM
 

Search in Used Dept. 



Mirek Elsner
Offline
• • •
Upload & Sell: Off
p.2 #11 · p.2 #11 · Bayer masks and DxO color depth measurements


Thanks for the insight, theSuede. It has been very helpful, but I still can't connect the dots. Basically, I am trying to understand if the claims that manufacturer XYZ sacrifices high iso performance to achieve better color at base iso are valid or not and if I can empirically verify them (by looking at DxO data or otherwise).

To get a very small noise covariance, you need low noise and good color separation on the CFA.

So, the DxO color depth measures combined effect of noise and CFA color separation. So if a sensor has extremely low noise and mediocre color separation, it can still score high. Am I on the right track?

MDFB's often have very good (to good!) color separation, but often not that brilliant DR - and their over-the-top color separation needs tempering to conform to human vision. But when you combine very good color separation with very low mid-gray noise levels (something MDFB's are extremely good at, but isn't directly related to DR!), you get really impressive gradation possibilities. It may not be perfectly translated into human-vision color, but it does look impressive.

If I want to see this color separation, what should I look at?

Thanks for your time!



Mar 14, 2013 at 03:21 PM
edwardkaraa
Offline
• • • • •
Upload & Sell: Off
p.2 #12 · p.2 #12 · Bayer masks and DxO color depth measurements


Canon by it's own admission used lighter CFA on the 5D2 to get better high iso performance. The comment was removed the next day of course. Like Philip said, the best judge is our own eyes. After all that's what photography's about. In my own experience cameras with better color separation can behave strangely in some lighting situations, while the others will produce quite ok results under all lightings but no wow factor. The A900 and M9 come to mind, and I have used both.


Mar 14, 2013 at 03:52 PM
Mescalamba
Offline
• • • •
Upload & Sell: Off
p.2 #13 · p.2 #13 · Bayer masks and DxO color depth measurements


edwardkaraa wrote:
Canon by it's own admission used lighter CFA on the 5D2 to get better high iso performance. The comment was removed the next day of course. Like Philip said, the best judge is our own eyes. After all that's what photography's about. In my own experience cameras with better color separation can behave strangely in some lighting situations, while the others will produce quite ok results under all lightings but no wow factor. The A900 and M9 come to mind, and I have used both.


Well, A900 has SMI 87/82 .. maybe that 5 point difference means real difference?



Mar 14, 2013 at 06:53 PM
theSuede
Offline
• • • •
Upload & Sell: Off
p.2 #14 · p.2 #14 · Bayer masks and DxO color depth measurements


Yes, when using more translucent filters (less light filtered away) you get better noise performance, since there's going to be more light for the sensor to measure. This has to be weighed against the noise increase you get from increased color processing.

The balance between them depends on how well you can mix & match with the available pigments that you can use in a CFA filter (not that many, if you're staying within reasonable costs), AND where you want the camera performance relative peak. A certain type of solution might increase high ISO performance, another might impair it. Both of them are quite probably available in both metamerically "good" versions and metamerically "bad" versions.

But if you can make the filters more translucent while still keeping room for spot corrections with low metamerical errors, you're basically still in good shape. Most color out in reality has a larger support area than we think. Raw converters now are mostly also able to follow shape clues, so that a red line is averaged for raw values (and hence color) ALONG the line, not across it (which would give very low saturation, aliasing and also luminance errors).

But the previously mentioned high-SMI Sony cameras have a rather large head start in some areas - since they can be linearly converted into something very similar to human eye SML response, they will inherently have less metamerical errors and also follow the "human" way to white-balance a scene more closely.

The most simple way to test is actually to have as many individually unique pigments as possible in one scene, under one light.

You can then get an average SMI value, the median error AND the average "worst case" errors. And if you look at colors close to each other, you can also see if the correction needed is smooth, or very fine grained.

What you want is a high SMI (but not necessarily the highest, there's always practical trade-offs), you want the average and median error to be the same, few "worst case" errors and a smooth spot correction map that doesn't have any kind of sudden spikes or valleys in very narrow bands of color.

When you can satisfy all of the above, you're good. What I'd like is the A900 filters on a D800/D600 body...
Canon's own little "solution" is good for some things - but bad in others. But as long as you can live with that color noise reduction HAS to be pretty strong - it's still pretty good after you've added the right correction (camera profile).

There's actually way more difference in the built-in camera jpg engines than there are in the actual CFA filters between most brands.



Mar 14, 2013 at 08:31 PM
theSuede
Offline
• • • •
Upload & Sell: Off
p.2 #15 · p.2 #15 · Bayer masks and DxO color depth measurements


carstenw wrote:
This is quite interesting to me. So what measurement can be made which takes these spot colour corrections into account, and thus gives an accurate real-world idea of how accurate the colours of a given camera are? Does DxO or some other site list such numbers?


DxO does those kind of tests, but the more detailed results are paying-customer-only. And they're not exactly cheap.

Basically what you need is a good color chart (or even better a known collection of materials like the duPont tile sets combined with a good variety of known spectrum other pigment prints - and plastics). Then you need to look at how smooth the spot correction is, after the initial linear correction by matrix. I've got a large collection of this kind of data, but most of that material is also client-owned.

To do this "on the cheap", a simple dE measurement on a color-chart (where you have a resolution-constrained color profile) does the job to a reasonable accuracy. If you do it in three lights - like sun, halogen and some known fluorescent bulb - you get very accurate data.

But I don't know if this is available online for free.



Mar 14, 2013 at 10:22 PM
carstenw
Offline
• • • • • •
Upload & Sell: Off
p.2 #16 · p.2 #16 · Bayer masks and DxO color depth measurements


With "chart" I presume you mean something like the Digital Color Checker, rather than the standard 24-patch? How hard is this to do, and can you give me a hint how good my D800 can get with a good profile? I am very happy with most aspects of this camera, but I have the feeling that a better profile could unlock some potential in the colour area.

Thanks for any help!



Mar 14, 2013 at 11:05 PM
Mescalamba
Offline
• • • •
Upload & Sell: Off
p.2 #17 · p.2 #17 · Bayer masks and DxO color depth measurements


Even Classic ColorChecker from X-rite with Adobe profile editor can help I think..

http://xritephoto.com/ph_product_overview.aspx?id=938&catid=28&action=overview

This might be probably better. But it needs software.. Which I think is sorta hyper-expensive. Dunno if there is some freeware one like Adobe profiler.



Mar 15, 2013 at 01:44 AM
Taylor Sherman
Offline
• • •
Upload & Sell: Off
p.2 #18 · p.2 #18 · Bayer masks and DxO color depth measurements


Very interesting stuff, thanks everyone.

Color sensing/processing gets more complicated the more I learn about it! The engineer in me feels pretty put off by it, really. It's such a "fuzzy" process. Audio processing, for one example, is so much simpler it seems (and it's even substantially more complex than many people realize, given the importance of phase response).





Mar 15, 2013 at 01:59 AM
carstenw
Offline
• • • • • •
Upload & Sell: Off
p.2 #19 · p.2 #19 · Bayer masks and DxO color depth measurements


Mescalamba wrote:
Even Classic ColorChecker from X-rite with Adobe profile editor can help I think..

http://xritephoto.com/ph_product_overview.aspx?id=938&catid=28&action=overview

This might be probably better. But it needs software.. Which I think is sorta hyper-expensive. Dunno if there is some freeware one like Adobe profiler.


That link is the digital version I was referring to.



Mar 15, 2013 at 10:00 AM
Mescalamba
Offline
• • • •
Upload & Sell: Off
p.2 #20 · p.2 #20 · Bayer masks and DxO color depth measurements


Ah yea, missed that digital part. I guess its not hard, just expensive. :/


Mar 15, 2013 at 01:27 PM
1      
2
       3       end




FM Forums | Alternative Gear & Lenses | Join Upload & Sell

1      
2
       3       end
    
 

You are not logged in. Login or Register

Username   Password    Reset password