7D Maze/low ISO artifacts UPDATE

This is certainly a possibility, though it seems a bit silly to do it this way. Canon would have to use this information in their jpeg engine and apply a correction prior to demosaic. Since they already have the corrected data in hand, it makes no sense not to write that data to raw, rather than write both the uncorrected data and the data that would be needed to apply the correction.
The more I think about it it may be something in those discarded rows and lines on the sensor. IMHO one could possibly extract the necessary configuration from those as well - and at the same time account for any changes in behaviour due to temperature or exposure length.
--
That might be, I haven't looked in detail at the structure of those pixels; I thought they were all black, though. If there were some that were given a fixed offset, however, and then sent through the amplifier/ADC, they would give an indication of the gain mismatches. Black pixels don't however, since there is no signal to apply the gain to.

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/
 
Could you send me the RAW files?

There is much less patterning on your examples than on BB's 7D (note that what I posted at FM had only 2EV brightening of levels).

It's curious that the 20D shows much less variation from patch to patch, while the patch boundaries are much more apparent on the 7D. It may be that this is a design feature, but I've seen 7D's which don't show much green channels variation at all.
--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/
 
Why I didn't see any samples on real world photography from 7D owners.

While on 5D2 so many real world samples
http://forums.dpreview.com/forums/read.asp?forum=1032&message=33189709
http://forums.dpreview.com/forums/read.asp?forum=1019&message=33396477

All I saw is blackframe analysis on 7D, but no real world samples

I doubt it is a common issue of 7D
The 'real world' 7D crop in the link I posted above has very visible maze artifacts in what seems to be an ISO 200, ACR image.
 
2. since they have internal ways to try to fight against this issue, they decided that they could care less and won't even try for 7D and future models, use DPP and live with how DPP gets around it or don't complain, lets pass cost of better QC on to third party developers to rewrite how they do everything and try to counter this
Maybe the whole thing you perceive as a problem is as intended and the odd one out which you thought is ok is the odd one out that misbehaves?

Give it a rest already. A beta version of ACR bears no impact on real life situations. I had hoped that the guys at adobe were less stupid by not adding a shoddy preliminary support for the camera. Had they not done this, you and all the others that you have put up to this kind of silly testing would be happy and be shooting real pictures instead of going through testing cycle after testing cycle without end but only proving how bad the ACR converter currently is!
--
I don't see why the onus is on software to deal with a hardware problem. dcraw exhibits the problem too. There is a kludge whereby one averages the green channels before interpolation, and this may be what DPP is doing too. So yes, at the cost of smearing the resolution the problem can be mitigated. So yes, everybody who has the problem should be happy, don't worry, as long as they turn their 18MP camera into a 10MP camera in terms of resolution the problem can be avoided.
This sort of inbalances has been in the digital sensor systems from day one - more or less visible, and once found they have been corrected in SW without any too big hassle. Should Canon has tweaked this in their RAW data as some people has proposed - I do not necessarily agree as this may limit what some better algortihms in a RAW SW can do. Or is it enough Canon handled this properly in their in-camera FW for JPEGs and in their DPP, and assume all the other RAW tool manufacturers do the same in their tools.

The step from this sort of filtering to full pledge Bayer domain noise etc. filtering may be smaller than we think. I'd like to delay this "raw data purification" as long as possible - eventually it will come as did the NR in the JPEGs as few years ago even though Canon evidently tried to resists the main tide.
 
Why I didn't see any samples on real world photography from 7D owners.
Why don't you see a lot of sheep with burnt noses from electric fences? Not because the electricity isn't there, but because the sheep stay away. People get burnt with using deep shadows, and they'll avoid them, even though it might be advantageous to use them, if possible. Many people whine that digital doesn't have the highlight headroom of negative film, but it actually does, because ISO 100 on a digital would be ISO 800 or more with film; digital is that much more sensitive. But, since the shadows of low ISOs have been marred by read noise and pattern artifacts, digital cameras have underrated their ISOs (IOW, exposed to the right) to stay clear above the read noise.

Without the patterns, one could shoot ISO 100 at -3 or -4 EV, or set the camera to "sunny f/16" on a bright, sunny day with deep shadow areas, and not worry about too much ugly noise. The first real image I subtracted the vertical offset patterns from, I examined in a shadow area pushed from ISO 100 to about 5,000, and the result didn't look very bad at all. And actually, I didn't subtract the pattern but an untouched blackframe, so I actually increased random noise by 1/2 stop, and it still didn't look too bad. A better subtraction image could be made. I will be demonstrating all this soon, but I have a lot of "real world" shooting to do, despite the caricature often made of me from people extrapolating my personality and activities from the facet of me seen in these forums.
All I saw is blackframe analysis on 7D, but no real world samples
I doubt it is a common issue of 7D
Of course it isn't, because the sheep are well-trained, and limit their possibilities by avoiding the issue.

But, relative to the 5D2, the 7D issue is more minor (on a good copy). The 5D2 banding has a lot of bold, single lines, and that causes color clustering or outright colored lines.

The DR of the 5D2 is more limited than the 7D at base ISO, from what I've seen, but I haven't done a direct comparison yet.

Anyway, I have run into the finiteness of 7D high-ISO capability; in dark green areas, (filiage itself in light filtered and reflected by foliage) horizontal banding is visible at 3200 and above. That started at about 1250 with the 50D, though.

--
John

 
Here is the difference of the two channels (CFA split using IRIS, difference viewed in ImageJ), the selection is on the cyan patch of your RAW:



The lower left white patch on the CC chart shows this in its difference:



The vertical streaking is a cause of interpolation artifacts, as is an overall shift of means which is most predominant in cyans and yellows. While the cyan patch #18 has a difference in means of -18 in G2-G1, the yellow patch #16 on the CC chart shows a difference in means of +38 for G2-G1. I believe this is statistically significant, even though it is several times smaller than the std devs of the patches, since the white lower left patch has a much higher std dev but a much smaller mean difference. So either the CFA is misaligned on your camera, or indeed as some have suggested Canon has decided to use a four-color CFA, with the two greens having somewhat different spectral responses. I hope to find the time to dig into this latter possibility later, by looking at whether the spectral responses of your CC chart agree with those of bronxbomber's copy, and the test images at IR.

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/
 
The 'real world' 7D crop in the link I posted above has very visible maze artifacts in what seems to be an ISO 200, ACR image.
Not an argument, it's ACR again. If you went to a doctor and told him: "It hurts when I hit myself really hard" he will only tell you to stop hitting yourself.

It's the same with this argument over and over again. ACR is not yet capable of converting 7D images. Period. Don't judge a camera on the quality of beta software with known limitations by a third party manufacturer!

--
regards
Karl Günter Wünsch
Visit my gallery at
http://www.fotocommunity.de/pc/pc/mypics/461808
 
Rawanalyze (RGB avg,stddev)

G2 4259, 50.6
G1 4307, 38.8

G1 4276, 47.2
G2 4228, 56.5

G2 6085, 67.4
G1 6116, 52.1

(or did they really do this on purpose and the good copy was the bad one???)

(very interested to hear what sort of numbers John gets from his new 7D in Rawanalyze using shift 5 and 4 on a patch and channel mode.)
I'm afraid I don't know how to use Rawanalyze and have no idea what these numbers mean. Could you explain them for me? How can I test mine? I have the program, running in VMware.

I'm sorry to see you're still having issues.

--
http://models.stevemelvin.com
 
The DR of the 5D2 is more limited than the 7D at base ISO, from what I've seen, but I haven't done a direct comparison yet.
Not what I've found.

In fact the DR of the 7D at ISO 100 is less than that of the 7D at ISO 200. There is definitely something odd at ISO 100 on the 7D, check out a set of files in RawAnalyze and take a look at the white points. ISO 100 is 13600 and ISO 200 is 15340. In side by side shots under the same lighting, the ISO 100 shots clip before the ISO 200 shots.
 
There does seem to be a consistent frequency response difference of the two green channels. I converted 3 separate images containing a color chart -- julio's, bronxbomber's, and the one at Imaging-Resource -- using dcraw in both 3-color and 4-color mode. I need to go back and check the code of dcraw, but I believe that all it does in 4-color mode is to average the two greens before interpolating. I then overlaid the two interpolations in Photoshop, using difference blending mode, and raised the levels 4 stops. Here are the results:

IR:



Julio:



bronxbomber:



There does seem to be a systematic shift between the two interpolations in the yellow and cyan patches of the color chart. I am back to thinking that Canon did use two different spectral responses for the two green channels.

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/
 
The DR of the 5D2 is more limited than the 7D at base ISO, from what I've seen, but I haven't done a direct comparison yet.
Not what I've found.

In fact the DR of the 7D at ISO 100 is less than that of the 7D at ISO 200. There is definitely something odd at ISO 100 on the 7D, check out a set of files in RawAnalyze and take a look at the white points. ISO 100 is 13600 and ISO 200 is 15340. In side by side shots under the same lighting, the ISO 100 shots clip before the ISO 200 shots.
I think this is nothing new. If memory serves me correctly, while ISO 200 has more noise, it tends to have better DR, and colour. It seems though, the 7D has a greater difference than should be acceptable.

--

 
The 'real world' 7D crop in the link I posted above has very visible maze artifacts in what seems to be an ISO 200, ACR image.
It is very important that people realize that there are two completely different issues going on here. When I see "low ISO" issues being discussed, I tend to think of offset banding, and that is how I replied. It would save a lot of confusing if people made it very clear whether they are talking about "mazing" or (blackpoint) "offset banding".

--
John

 
In fact the DR of the 7D at ISO 100 is less than that of the 7D at ISO 200. There is definitely something odd at ISO 100 on the 7D, check out a set of files in RawAnalyze and take a look at the white points. ISO 100 is 13600 and ISO 200 is 15340
The clipping levels are in fact different, but that does not mean, that ISO 100 "clips sooner". Rather, the ISO 100 pixel values should be seen "more compressed"; this is not in terms of storage occupation, but it terms of the "resolution of the dynamic range".

Although the camera is said to support 14bit depth, that is not true. The effective pixel values start at around 2048, and the clipping occurs at 12650, 13580 or 15300, approximately, depending on the ISO; however, the lower number of possible pixel values, i.e. the lower number of tonal levels do not indicate a lower dynamic range.

In fact, the 7D's DR is greater with ISO 100 than with 200, though I would not put any number on that for lack of suitable raw images.

--
Gabor

http://www.panopeeper.com/panorama/pano.htm
 
The DR of the 5D2 is more limited than the 7D at base ISO, from what I've seen, but I haven't done a direct comparison yet.
Not what I've found.

In fact the DR of the 7D at ISO 100 is less than that of the 7D at ISO 200. There is definitely something odd at ISO 100 on the 7D, check out a set of files in RawAnalyze and take a look at the white points. ISO 100 is 13600 and ISO 200 is 15340. In side by side shots under the same lighting, the ISO 100 shots clip before the ISO 200 shots.
I didn't check the whitepoint differences between 100 and the rest on the 7D, but even with the lower clipping point of ISO 100, the standard deviation of read noise is lower on the 7D by about the same ratio as 5D2 max signal to 7D max signal at ISO 100, and with the periodic element removed from the 7D, the practical noise floor should that much lower, especially with images displayed at a small size.

The vertical offset banding in the 5D2 is more random and unique to each frame, it seems.

--
John

 
Yeah it does make one wonder about the differences since they seem to be so common. It's starting to look like g1g2 differences are there on most (all?) 7Ds to a greater or lesser extent. That would seem to indicate either a general design flaw of the filters or deliberate design.

I'm also seeing noticeable small visual difference when comparing the standard and 4 color Dcraw tiffs in CS2 which are confirmed by the histograms when I switch the focus of the two back and forth quickly. The standard 3-color version shows more saturated colors pretty well across the board and its histogram is shifted upwards slightly relative to the 4 color conversion. I also added a color sampler at the same coordinates in both tiffs on the yellow patch. The standard conversion has an RGB reading of 154, 95, 0 and on the 4-color conversion it's 160, 96, 0.

Is that just a normal byproduct of the 2 different conversions visible with any camera or am I seeing a visual clue of a global effect of the 7D filter that increases color saturation if processed by normal means?

Another odd thing is that this difference is easy to see in PS CS2 but not at all if I look at those tiffs in LR 2.5 or LR 3.0.

--
http://fotoman99.smugmug.com/
 
There does seem to be a consistent frequency response difference of the two green channels. I converted 3 separate images containing a color chart -- julio's, bronxbomber's, and the one at Imaging-Resource -- using dcraw in both 3-color and 4-color mode. I need to go back and check the code of dcraw, but I believe that all it does in 4-color mode is to average the two greens before interpolating. I then overlaid the two interpolations in Photoshop, using difference blending mode, and raised the levels 4 stops. Here are the results:
There does seem to be a systematic shift between the two interpolations in the yellow and cyan patches of the color chart. I am back to thinking that Canon did use two different spectral responses for the two green channels.

--
emil
So far all of the (perhaps) bad copies have shown G1 gain greater than G2 gain on the patches I looked at, however my samples size was too small to be significant, since one commonly sees a number of heads or tails in a row when flipping a coin, but it certainly would fit with something new going on (or a universally off production run).

Maybe I should have taken my first copy back to Canon instead of the store....

I will try to post up a new colorcheck from my new copy later.

One thing though, what about John who claims that his greens are the same? And din't you say you saw another copy with equal greens? And why is Julio's much closer than my first copy?

Also, if 4-color tends to rob resolution if you split greens, might not one either be stuck with maze artifacts and harshness and extra noise or something less than 50D resolution while taking up 18MP of storage space? Or maybe they use adaptive algorithms picking out smoother swatches for 4-color and less smooth swatches for 3-color, sounds like a little bit of a mess and not necessarily something that third party converters would do, at least not for a long while, so one might end up stuck with DPP??
 
Rawanalyze (RGB avg,stddev)

G2 4259, 50.6
G1 4307, 38.8

G1 4276, 47.2
G2 4228, 56.5

G2 6085, 67.4
G1 6116, 52.1

(or did they really do this on purpose and the good copy was the bad one???)

(very interested to hear what sort of numbers John gets from his new 7D in Rawanalyze using shift 5 and 4 on a patch and channel mode.)
I'm afraid I don't know how to use Rawanalyze and have no idea what these numbers mean. Could you explain them for me? How can I test mine? I have the program, running in VMware.

I'm sorry to see you're still having issues.

--
http://models.stevemelvin.com
i measured smooth even patches and the first numbers are the avg raw values recorded for a green channel over part of that path and the second number tells you how basically how much the value is varying from photosite to photosite for that green channel in the sampled part of the patch.

But this whole thing appears to be getting confusing again, so I'm not sure what the story is now:

1. canon did swith to 4-color filters (so the bad copies are actually 'bad' i.e. good), but this would seem likely to cause some issues that might be tricky to get around and I wonder if it might not be a long, long time before one could use a third party converter and get the most of the 7D then (???) and one might not be stuck with either mazing or more like something less than 50D resolution maybe even close to 12MP sensor in the meantime or always stuck with some tricky adaptive algorithm likely to not ever handle all parts of an image quite as well as a 3-color one could (??? I have to confess that while I once almost started looking into converter processing I ended up not, so I really don't have much technical knowledge about that sort of thing at all at the moment)

2. canon has some sort of problem on the assembly line where they are making the gain higher on one of the greens on some (many it seems, most even?)
copies so far

3. canon has some sort of problem where many of the CFA filter arrays have been sloppily aligned to the sensors, tending to be pushed off to one side to one degree or another

4. something else
 

Keyboard shortcuts

Back
Top