Matt Grum wrote:
Potentially, yes, if you had a very low DR scene and stretched it out. But in general I'd much rather have more resolution in the shadows.
I highly suspect this is what should be tested - and perhaps is more practically relevant to a discussion of bit depth- as it's not too uncommon to take a lower contrast, normally exposed scene (or even local area in a scene) and stretch it out, increase micro contrast and so forth. So for a test of bit depth - whether it's 11/12 bit vs 14bit, 16bit or even the effect of Sony's raw compression scheme - maybe folks should shoot a relatively flat scene with subtle detail and then expand the contrast. This should show what is going on in subtle gradations at various bit depths. We should likely see banding, posterization and so fourth in the lower bit file first when the files are really pushed/ stretched out as there is of course less information there.
I would really like a definition of "show" when talking about effects of compression. Most discussions I've read are concerned only with gross artifacts: clusters of pixels that clearly should not be in the image.
I've not seen discussions of how compression might create minor shifts that degrade the IQ but are not immediately obvious. Hopefully this will become moot soon, but these analyses help bring that moment closer.
Davide Roveri wrote:
This is a great post! Following up on the quick test i done last night i pushed the Exposure to +5EV and the difference between the mechanical shutter (on the left hand side) and the electronic one still is remarkably small.. I don't have the technical knowledge to know what they have done in order to obtain this results but as a user i'm seriously impressed.. :-)
Joel Cahane wrote:
Small difference ?
Seems to me the letters (north, south etc...) on top are entirely missing on the image taken with the electronic shutter.
That's because the clouds have moved and the letters aren't showing up against the brighter background - nothing to do with the electronic shutter!
Tariq Gibran wrote:
I highly suspect this is what should be tested - and perhaps is more practically relevant to a discussion of bit depth- as it's not too uncommon to take a lower contrast, normally exposed scene (or even local area in a scene) and stretch it out, increase micro contrast and so forth. So for a test of bit depth - whether it's 11/12 bit vs 14bit, 16bit or even the effect of Sony's raw compression scheme - maybe folks should shoot a relatively flat scene with subtle detail and then expand the contrast. This should show what is going on in subtle gradations at various bit depths. We should likely see banding, posterization and so fourth in the lower bit file first when the files are really pushed/ stretched out as there is of course less information there. ...Show more →
Sony's 12 vs 14 bit mode is not just an encoding difference but a difference in how the ADC is operating, based on the extra noise visible in the images that is not present in other cameras' 12 vs 14 bit modes like Nikon's. That extra noise will dither any tonality difference so none will be visible separate from the noise.
Regarding Sony's 11+7 compression affecting tonality, it'll nearly impossible to discern outside of the already reported artifacts from sharp gradations and those too require extreme shadow pushing to see except in some specific star trail scenarios. Here are two of my examples from the compression thread last year, which also include Nikon 12 vs 14 bit comparisons: here and here).
Matt Grum wrote:
That's because the clouds have moved and the letters aren't showing up against the brighter background - nothing to do with the electronic shutter!
Davide Roveri wrote:
This is a great post! Following up on the quick test i done last night i pushed the Exposure to +5EV and the difference between the mechanical shutter (on the left hand side) and the electronic one still is remarkably small.. I don't have the technical knowledge to know what they have done in order to obtain this results but as a user i'm seriously impressed.. :-)
Matt Grum wrote:
That's because the clouds have moved and the letters aren't showing up against the brighter background - nothing to do with the electronic shutter!
But they aren't just hard to see — they are completely gone! We can't rule out an absolutely perfect match between color and luminosity of the letters and clouds, but we also can't rule out some sort of compression scheme that discards smaller data differences in a local area. (Note also that the section of the top vertical extension of the tower that is the same tonality as the missing letters is still visible. Odd.)
I think you might want to reshoot this one... just to be sure.
Conceptually, if you move from a 14 bit representation to a 12 bit representation, there are a few more or less hard facts about the data result. Either you decrease the dynamic range and keep the increment of difference represented by each step the same or you cover the same dynamic range and make the steps larger. There really is no other alternative.
There are a few ways you can deal with this reality. You can offset the range that is represented by the 12 bits, giving up the low end or the high end or perhaps a bit of both. You can alter the response curve to allocate more bits in the regions where you think more precision is necessary. You could expand the 12 bit representation and attempt to "fill in" the missing data via some kind of predictive process.
If you are looking to understand the result you would want to test a sample that contains the full dynamic range with lots of intermediate gradations included.
(In the end, it is quite possible that a) there will be differences but b) they will not be significant in many cases.)
gdanmitchell wrote:
But they aren't just hard to see — they are completely gone! We can't rule out an absolutely perfect match between color and luminosity of the letters and clouds
Of course we can't rule it out it's the only explanation that makes any sense.
gdanmitchell wrote:
we also can't rule out some sort of compression scheme that discards smaller data differences in a local area. (Note also that the section of the top vertical extension of the tower that is the same tonality as the missing letters is still visible. Odd.)
That's easy to rule out, given a) the details of the RAW compression scheme are all known, and b) the noise has been preserved, so something that could remove the lettering whilst completely preserving the noise (which also counts as "smaller data differences" seems completely impossible.
gdanmitchell wrote:
(Note also that the section of the top vertical extension of the tower that is the same tonality as the missing letters is still visible. Odd.)
That section of the top vertical extension is clearly brighter in the first image, so it's not odd at all that it shows up against the brighter background in the second image.
gdanmitchell wrote:
Conceptually, if you move from a 14 bit representation to a 12 bit representation, there are a few more or less hard facts about the data result. Either you decrease the dynamic range and keep the increment of difference represented by each step the same or you cover the same dynamic range and make the steps larger. There really is no other alternative.
Far too much is being made of the difference in bit depth, the differences we're seeing in the images (increased shadow noise, reduced dynamic range) are greater than you would expect from what is in reality (due to later lossy compression) a truncation of 1 bit in the source data.
There is clearly something else going on in the ADC in these modes.
snapsy wrote:
I'm not seeing any difference in those. I wonder if bulb @ 30 seconds is really using 12-bits.
My speculation: perhaps the camera detects that BULB @ 30 seconds was used, so auto-magically changes it to Manual @ 30 seconds.
If BULB @ 31 or 35 seconds can be set, that might force the camera to 12-bit mode for a better test.
You'd compare Manual @ 30 sec to BULB @ 31 seconds (each at base ISO).
ldallan wrote:
My speculation: perhaps the camera detects that BULB @ 30 seconds was used, so auto-magically changes it to Manual @ 30 seconds.
If BULB @ 31 or 35 seconds can be set, that might force the camera to 12-bit mode for a better test.
You'd compare Manual @ 30 sec to BULB @ 31 seconds (each at base ISO).
Or not?
Based on my A7s tests I think it's because Fred's images were only pushed +4EV. The differences are small at +4EV but much more obvious at +7EV, which of course argues that the differences aren't that significant for most uses.
Matt Grum wrote:
That's easy to rule out, given a) the details of the RAW compression scheme are all known, and b) the noise has been preserved, so something that could remove the lettering whilst completely preserving the noise (which also counts as "smaller data differences" seems completely impossible.
Not necessarily, depending on the nature of software involved. I have no idea — which is why another test is in order — what might be going on here, but it isn't hard to conceive of software that discounts the significance of luminosity data differences that likely existed between the letters and the clouds but does not discount the significance of a similar difference that is part of a linear object between brighter and darker points.
Yes, that is a supposition, but it is based on some knowledge of how certain kinds of data compression are accomplished, and they can occasionally produce odd and unexpected results. (Those who know how the mp3 audio compression scheme works may get what I'm saying here.)
That section of the top vertical extension is clearly brighter in the first image, so it's not odd at all that it shows up against the brighter background in the second image.
But note that this section contains a tonal gradient that ranges from brighter than to darker than the background sky, including tones that are identical to the sky luminosity. Yet here those middle tones still appear...
Far too much is being made of the difference in bit depth, the differences we're seeing in the images (increased shadow noise, reduced dynamic range) are greater than you would expect from what is in reality (due to later lossy compression) a truncation of 1 bit in the source data.
I acknowledged the possibility (probability, even) that there may be differences but the differences may not be significant. Trust me, that is a concept I'm well aware of and one that I've had to defend all too often around here! (As a printmaker I frequently shake my head at some of the things that people imagine will make critical differences in the final quality of their photographs...)
There is clearly something else going on in the ADC in these modes.
Indeed! I'd like to see this particular test run again, but with a more consistent test target. There is a good chance that such a test would confirm that the real-world differences are generally insignificant. But it might also raise a few more questions about the mysteriously missing letters on the weather vane.
Fred Miranda wrote:
The 12-bits Bulb limitation is indeed a big deal for landscape photography.
Agree. One of the reasons I've stayed with Canon (so far) is the availability of Magic Lantern enhancements. (but a7Rii is pre-ordered)
So what if Canon decided to cripple the 6d with no intervalometer for example (strictly firmware)? ML added it. Same for advanced Bulb and much more like Auto-ETTR, RAW histograms/blinkies/review, Dual-ISO, Auto-Micro-Focus-Adjustment-Dot-Tune, and so on (sorry for the ML fanboy enthusiasm).
My point is that I speculate a ML-type team for Sony could get around such limitations ... without causing the camera to brick.
Out of curiosity: my understanding is that the Canon firmware architecture allows ML to augment the "native firmware". Is such a thing possible with Sony?
I really, really wish several of the ML team would "fork" over to Sony and tackle the a7x family.
I sometimes wonder if Sony's R&D team has these discussions. And if they do, why they made the choices we have to deal/work with. It would help a lot if we could get real answers instead of guessing.
I think that this issue is likely to be much more apparent with B&W shooting rather then color work, so I think it is a relevant issue to be aware of.
B&W has a few characteristics that may cause the problem to be more apparent, including:
1. You have to 'throw out' a bunch of the sensor data to convert to B&W if you use a channel mixer (filtration) approach to conversion, which is the best way to get something that has the tones in the image that you might desire. If you just convert to B&W, you don't lose the majority of the channel information, but you have no control over the tonal distribution.
2. B&W shooters will often do much more manipulation of an image than you will typically see in color. The only things that I see coming close are the heavily overcooked HDR images that some people seem to like (you know, the ones that make your eyes hurt).
3. (At least in my case), I seem to heavily rely on orange or red filtration to get the tonal distribution in the image that I want. It seems that the R channel is prone to noise much more so than G to be sure (not a surprise because there are 2x as many G pixels in a Bayer array, and the filtration is somewhat less to achieve the G as well).
Anyway, I'm very interested in seeing how the new a7rII performs in tests like this. I haven't looked closely at this point WRT B&W images, but I was beginning to wonder if the increasing DR of the sensor is to the point where Sony needs to begin producing 16-bit sensors because even greater manipulation of the scene will be occurring to achieve a 'normal' looking image.
B&W shooters will happily take the wider DR in the sensor, but I think that it will mean that even more curves adjustments may occur in post than before, and the same goes for color work to a certain extent. What occurs is the midtones start becoming thin as a greater percentage of the 14- (or 12-) bit files fall in the highlight/shadow regions.
Note that this is occurring, because the samples that I have seen and people have commented on that have come SOOC have appeared flat to people. Getting the snap back into those images spreads out the midtones somewhat.
ldallan wrote:
Agree. One of the reasons I've stayed with Canon (so far) is the availability of Magic Lantern enhancements. (but a7Rii is pre-ordered)
So what if Canon decided to cripple the 6d with no intervalometer for example (strictly firmware)? ML added it. Same for advanced Bulb and much more like Auto-ETTR, RAW histograms/blinkies/review, Dual-ISO, Auto-Micro-Focus-Adjustment-Dot-Tune, and so on (sorry for the ML fanboy enthusiasm).
My point is that I speculate a ML-type team for Sony could get around such limitations ... without causing the camera to brick.
Out of curiosity: my understanding is that the Canon firmware architecture allows ML to augment the "native firmware". Is such a thing possible with Sony?
I really, really wish several of the ML team would "fork" over to Sony and tackle the a7x family. ...Show more →
Yes, it is entirely possible since Sony has its appstore and one user found a way to create custom apps (No -- not custom WIFI-apps, custom native apps). Via such a custom app one could load an entire interface. It would be quite easy since one also can load the app via a Custom Button and the "Application list", no need to have some special card with the UI etc.
What's missing is "only" people that understand Android development and that are willing to do the work (yes -- Sony firmware uses Android).
mjm6 wrote:
I think that this issue is likely to be much more apparent with B&W shooting rather then color work, so I think it is a relevant issue to be aware of.
B&W has a few characteristics that may cause the problem to be more apparent, including:
1. You have to 'throw out' a bunch of the sensor data to convert to B&W if you use a channel mixer (filtration) approach to conversion, which is the best way to get something that has the tones in the image that you might desire. If you just convert to B&W, you don't lose the majority of the channel information, but you have no control over the tonal distribution.
2. B&W shooters will often do much more manipulation of an image than you will typically see in color. The only things that I see coming close are the heavily overcooked HDR images that some people seem to like (you know, the ones that make your eyes hurt).
3. (At least in my case), I seem to heavily rely on orange or red filtration to get the tonal distribution in the image that I want. It seems that the R channel is prone to noise much more so than G to be sure (not a surprise because there are 2x as many G pixels in a Bayer array, and the filtration is somewhat less to achieve the G as well).
Anyway, I'm very interested in seeing how the new a7rII performs in tests like this. I haven't looked closely at this point WRT B&W images, but I was beginning to wonder if the increasing DR of the sensor is to the point where Sony needs to begin producing 16-bit sensors because even greater manipulation of the scene will be occurring to achieve a 'normal' looking image.
B&W shooters will happily take the wider DR in the sensor, but I think that it will mean that even more curves adjustments may occur in post than before, and the same goes for color work to a certain extent. What occurs is the midtones start becoming thin as a greater percentage of the 14- (or 12-) bit files fall in the highlight/shadow regions.
Note that this is occurring, because the samples that I have seen and people have commented on that have come SOOC have appeared flat to people. Getting the snap back into those images spreads out the midtones somewhat.
therealthings wrote:
If my focus was mainly b&w photography i would have chosen a sigma dp quattro because of it's sensor design.
Interestingly, when I beta tested the Dp Quattro, I encountered some subtle banding issues in my B&W's (converted from raw) that I did not experience with the A7r shot and processed similarly side by side. This was pre-release so maybe it was fixed.
There are so many other things that can cause issues with banding and so forth (quality of electronics/ signal path, etc.) that it's tricky to point to say 12 vs 14 bit, sensor design - or any one variable - as the culprit. All that said, the most malleable files I have processed for B&W came from a Leaf medium format back which I believe was 16 bit. I expect that back also had good, low noise electronics as well.
Fred Miranda wrote:
True, I went through 3 adapters to find a centered one. (parallel mounting flanges)
Reminded me of the pain of buying a Samyang lens.
Fred,
I will be using a Metabones IV with Canon glass on a Sony A7R II. Can you elaborate on what you mean by a centered adapter? Are you talking about how flush it is against the camera body? Do you think the Sony A7R II will have the same issue using adapters?
Lastly, how does this decentering show up in the real world? What tests, specifically, can we do to test for decentering?
I don't want to hijack this thread so maybe I should start a new post. I, and I am sure others, would appreciate your feedback.
Here is a quick link to a Lensrentals post on how to check for a decentered lens. I assume that it would also apply with an adapter attached? Just examine first with the raw lens, then chaeck the image circle with the adapter to see how close they are.
I have something similar set up in my basement. I use a chart to test $1,000+ and $2,000+ lenses when I buy them. A Canon 24-70 2.8 II, Canon 70-200 2.8 II, Sigma 120-300 2.8 Sport, Nikon 24-70 2.8 and Nikon 70-200 VR II, etc. I had to send back one used Nikon 24-70 after comparing it to my Nikon 70-200 VR II at 70 mm.
I would have to re-read Roger's article to see what he says, but I like to take test shots using my studio strobe equipment. The flash helps to freeze any movement, so that I can don't have to go to mirror lock up, self timers, worry about bumping the tripod, passing trucks, etc.
