My point about DXO is that the pattern noise here is not only visibly disturbing, but the large difference in quantity is also evident in the histograms. The pattern noise is the dominant noise in the low ISO black frames, and the standard deviation will pick that up.
FashionBoy wrote:
Brainiac and anyone else, please send your findings and images to Chuck Westfall via the Digital Journalist website (I have his email also and can dig that up), and I will do the same.
I am sure this is a firmware issue. Prints I have made from the 5D2 sample images supplied by Canon do not show these deficits (pre-1.0.6 firmware), and neither do prints made with the 1.0.6 firmware.
I am convinced the 5D2 is on par with the 1DS3 otherwise, and this is not "normal" performance.
Way too much shadow noise at low ISO.
I could be wrong... but I think in the provided lenscap samples there were also FW 1.0.6 5D2's included.
OTOH I had a FW 1.0.6 5D2 that was also better regarding shadow noise than the ones I converted to FW 1.0.7... Could be a coincidence though.
If you look at the results of this poll, it is pretty clear that people are seeing this problem in various raw developers, not just DPP. I have just upgraded to DPP 3.6, but the problem still occurs. I am not about to spend time, effort, and money on alternate raw developers which may or may not work better than DPP now or in the future when the majority of evidence suggests that they are no better. Fiddling with noise suppression features is not likely to be a solution to this, because the patterned noise is on a scale that makes it indistinguishable from important detail. This is a kind of noise problem that can probably only be solved at the capture stage, i.e. through a firmware fix, or by having the camera modified/repaired.
So far we have no evidence whatsoever that there exist 5D2s which do not exhibit this problem regardless of firmware. If someone reckons she has a 5D2 without this problem, please do the lenscap test as described earlier in this thread and share the results with us. It would be very good news.
ejmartin wrote:
One possibility is that the data rates for 21MP are forcing a faster throughput for the amplifier and ADC, and that results in more RF noise that appears as pattern noise when the pixels are read out sequentially.
The 1Ds3 does not suffer from this, so it's not about 21 megapixels. It might be about having one DiG!C instead of two, but more likely it is about compromises in the camera design due either to price or the inclusion of video.
> As for why DXO data don't indicate the level of pattern noise, the contribution of pattern noise to the noise of any individual pixel is actually rather small, negligible in the standard deviation of pixel values relative to other noise sources that are uncorrelated from pixel to pixel. The reason it is so visually apparent even though its magnitude is small is that it is coherent across many pixels in a row or column, while the random noise averages out; our perception is designed to detect edges and patterns and so readily integrates the data across pixels to see the pattern. So DXO data and any other tester measuring only the standard deviation of pixel values will not pick up the pattern noise, even though it may be quite apparent to the viewer. There are other measures besides the standard deviation that can quantify the pattern noise, but I haven't done any analysis as yet....Show more →
That is very well put, and precisely the reason why I think DXOmark's graph-based appraisal of cameras is worse than useless. It's like appraising a car based on laboratory tests of the components of performance. One drive on the right road will tell you that despite its elevated measurables, the car is a coffin on wheels (like all US sportscars ;-).
brainiac wrote:
The 1Ds3 does not suffer from this, so it's not about 21 megapixels. It might be about having one DiG!C instead of two, but more likely it is about compromises in the camera design due either to price or the inclusion of video.
I did say it might be about the *data rates* for 21MP, and of course having only one processor instead of two (as I was well aware when I wrote the above) would affect that strongly.
That is very well put, and precisely the reason why I think DXOmark's graph-based appraisal of cameras is worse than useless.
They measure what they will, and I find that information quite valuable (having been in that game for some time). That they don't measure everything they could is certain, and probably related to what their RAW conversion software is designed to do. If they had a pattern noise reduction algorithm, they probably would be measuring pattern noise.
But I find these sorts of complaints a bit annoying -- we get vastly more information about sensor performance than we ever had, and people are still kvetching that this or that property (that is specific to their brand and even model of camera) is not included.
alundeb wrote:
My point about DXO is that the pattern noise here is not only visibly disturbing, but the large difference in quantity is also evident in the histograms. The pattern noise is the dominant noise in the low ISO black frames, and the standard deviation will pick that up.
How are you seeing it in the histograms?
I beg to differ that pattern noise is the dominant noise as measured by the standard deviation. Have a look at Figure 3 at
Mouse-over the image to see how much the fixed pattern noise is reduced by subtracting a pattern noise template from the RAW data. Nevertheless (as noted in the surrounding text) the standard deviation of the pixel values is reduced by only 20%.
Also note that people here are drawing conclusions based on the output of RAW converters, which may (depending on the design of their demosaicing algorithm) enhance the pattern noise if they think it represents edges that the interpolation should be following. It is much better if one is going to make quantitative claims to analyze the RAW data directly in a program like IRIS.
I'm getting very confused, could someone briefly summarize the position please. Is 50 ISO better than 100, are standard ISOs (100, 400, 800, 1600 etc) better than the intermediate ones on a 5D2?
ISO 50, 100, 125, as well as 200HTP and 250HTP are all derived from a hardware gain of ISO 100 by digital manipulation post-capture
ISO 160, 200, 250 as well as 320HTP, 400HTP and 500HTP are all derived from a hardware gain of ISO 200 by digital manipulation post-capture
ISO 320, 400, 500 as well as 640HTP, 800HTP and 1000HTP are all derived from a hardware gain of ISO 400 by digital manipulation post-capture
and so on. What is changing when you dial in these various ISO settings (and HTP) is the camera's metering relative to the actual hardware gain being employed -- the camera is either overexposing or underexposing the RAW image. So ISO 50 is better than ISO 100 if you have the highlight headroom to overexpose by a stop. ISO 250 is actually NOT better than ISO 200 -- the camera multiplies the quantized pixel values by 1.25, so you lose the top 1/3 stop of highlight data that you would retain if using ISO 200 at the same shutter speed and aperture.
Mouse-over the image to see how much the fixed pattern noise is reduced by subtracting a pattern noise template from the RAW data. Nevertheless (as noted in the surrounding text) the standard deviation of the pixel values is reduced by only 20%.
Also note that people here are drawing conclusions based on the output of RAW converters, which may (depending on the design of their demosaicing algorithm) enhance the pattern noise if they think it represents edges that the interpolation should be following. It is much better if one is going to make quantitative claims to analyze the RAW data directly in a program like IRIS....Show more →
Thanks, I have no problem with this, and sorry for not being specific, i meant pattern noise being dominant for the 5D2 samples here in particular.
Anyway, here are the ISO 100 histograms of two resized files, top: 1D3s , middle: 5D2, bottom: both overlaid. In the latter, the horisontal position is aligned to approximately the same mean value. The standard deviation relates to the width of the histogram. How is it possible that these histograms can be derived from RAW files with practically identical standard deviation, using the same raw converter? Is all this algorithms? What am I missing?
p.18 #11 · Are you seeing 5D2 Criss-Cross Pattern?
alundeb wrote:
Anyway, here are the ISO 100 histograms of two resized files, top: 1D3s , middle: 5D2, bottom: both overlaid. In the latter, the horisontal position is aligned to approximately the same mean value. The standard deviation relates to the width of the histogram. How is it possible that these histograms can be derived from RAW files with practically identical standard deviation, using the same raw converter? Is all this algorithms? What am I missing?
The effect of resampling an image on the standard deviation depends to some extent on the image content. For instance, pattern noise does not decrease upon resampling, because it is coherent across many pixels, while random noise does average out, being spatially incoherent. So to some extent what you are seeing (if as you wrote the files have been resized) is the fact that the 5D2 has more pattern noise than the 1Ds3. To see whether this is what's going on, try overlaying the two noise distributions without resizing the images.
There are other things that can make two different looking noise distributions have the same standard deviation, for instance a few outliers in the distribution. For instance hot pixels are way out on the tail of the distribution, but just a few of them can substantially increase the standard deviation of the whole collection (thus it's best to analyze a selection where there are none of them, rather than the entire image, if one is using a raw converter that doesn't eliminate hot pixels or if one is analyzing the raw data directly).
p.18 #12 · Are you seeing 5D2 Criss-Cross Pattern?
Good.
Here are the histograms of unscaled crops. The shape is different, but the relative difference between 1D3s (top) and 5D2 (bottom) is the same.
I don't see outliers in the 1Ds3 image that could increase its stddev to the same level as 5D2.
p.18 #15 · Are you seeing 5D2 Criss-Cross Pattern?
ejmartin wrote:
Are these black frames or underexposures? If the latter, what is the exposure for each image (Tv, Av, ISO)?
Usually black frames yield a somewhat Gaussian distribution in the raw data, so I'm puzzled by the shoulder in the 5D2 distribution.
These are the ISO100 black frames posted by brainiac and Daan B at page 5 or 6 in this thread. If you don't find all details on exposure and processing there, they may be able to fill in.
p.18 #18 · Are you seeing 5D2 Criss-Cross Pattern?
From Canon white paper on 5D2:
"... These optical advances and technology improvements to the sensor, plus countermeasures that help suppress noise from other camera circuits, are responsible for the sensor’s high signal to noise ratio, enabling higher ISO speeds and improved dynamic range at low ISOs."
p.18 #19 · Are you seeing 5D2 Criss-Cross Pattern?
alundeb wrote:
From Canon white paper on 5D2:
"... These optical advances and technology improvements to the sensor, plus countermeasures that help suppress noise from other camera circuits, are responsible for the sensor’s high signal to noise ratio, enabling higher ISO speeds and improved dynamic range at low ISOs."
p.18 #20 · Are you seeing 5D2 Criss-Cross Pattern?
Daan B wrote:
Compared to what?
Or is it just some hollow marketing phrase?
Compared to the 5D, and even without resampling the image.
The text continues:
"Those are impressive sensor improvements,
considering that the original EOS 5D had
larger 8.2ìm pixels compared to 6.4ìm pixels
on the EOS 5D Mark II, and that pixel
noise usually increases as size decreases."
I don't find this hollow, I still think there is a problem.
