Initial testing shows Canon’s new 32.5mp APS-C sensor improves dynamic range over predecessor

SteveC

M50 & T6i
Sep 3, 2019
352
202
Same here, I already have an 15-45 and the deal for the 18-150 isn't that good since it's basically full price body + full price lens + "free" EVF. The white box version of that lens is already a lot cheaper and looking at the grey market it's even less. So for a savvy shopper it's more "half price EVF" and "free EVF".
So I preordered the M6II body only and will see after a few weeks if I actually want the EVF. Hopefully there well be white box versions of it by then :)

(The M6II release date is also on day 3 of the family vacation abroad, so I'm slightly miffed about that)
I just replied to a comment that pointed out a BIG drawback of not having a viewfinder--the screen won't be worth beans in bright daylight. I remember once having no idea if I was actually taking a video or not--a Galapagos tortoise moving at warp speed [well by his standards, anyhow]. I literally couldn't see whether the red dot was "on" or not. (As it turned out, I wasn't taking the video when I wanted to be. :mad:) I was tempted to start packing a black T-shirt to serve as a cover, like the 19th century photographers used.
 

knight427

EOS 80D
Aug 27, 2018
104
160
Yes... 15 stop of DR in a 14 bit file..... Can you say cooked data?
I don’t know any details behind any of the DSP going on in cameras, but I do know a bit about perception versus measurement. Nearly all human perceptions are best approximated by logarithmic scales (this includes time, which is why we perceive time to be going faster as we get older, 1 year is proportionally a very small fraction of your life when you are 40, but a huge fraction when you are 8). Brightness is no exception. We perceive doublings of brightness to be a significant change. If we adopt an arbitrary unit of measurement, which can only be recorded in integers from 0-10,000, this means changing in brightness from 1 to 2 is as significant a change as 5,000 to 10,000. I’d be shocked if any camera manufacturer ignores this reality and records data linearly. My assumption has always been that DSP is used on every camera to make better use of the bits available. I have no idea if that assumption is reasonable.
 

dtaylor

Canon 5Ds
Jul 26, 2011
1,397
840
Basically you've provided a long explanation on why noise reduction works for Canon but doesn't work for Sony!
Go open two RAW files at two different ISOs in ACR. Set to 0 Color NR the lower ISO one should just show color noise. The higher ISO one should have more color noise. Now apply the same amount of CNR to both. Tell me which one benefits more.
 

Sharlin

EOS 6D MK II
Dec 26, 2015
1,060
557
Turku, Finland
I don’t know any details behind any of the DSP going on in cameras, but I do know a bit about perception versus measurement. Nearly all human perceptions are best approximated by logarithmic scales (this includes time, which is why we perceive time to be going faster as we get older, 1 year is proportionally a very small fraction of your life when you are 40, but a huge fraction when you are 8). Brightness is no exception. We perceive doublings of brightness to be a significant change. If we adopt an arbitrary unit of measurement, which can only be recorded in integers from 0-10,000, this means changing in brightness from 1 to 2 is as significant a change as 5,000 to 10,000. I’d be shocked if any camera manufacturer ignores this reality and records data linearly. My assumption has always been that DSP is used on every camera to make better use of the bits available. I have no idea if that assumption is reasonable.
RAW pixel values absolutely are linear, they’re literally electron counts with some gain and bias applied. The top EV of DR literally encompasses half the range of representable values of a RAW pixel. This is why ”expose to the right” is a thing—as long as you’re not saturating highlights, overexposing maximizes the information content of your RAW files.

It would make sense to compress the top EVs if you have deep photosites but are stuck with a lower bit depth digital side. But usually sensors are designed with the A/D conversion in mind, so that at native ISO full well capacity roughly corresponds to the maximum representable digital value after conversion.
 
  • Like
Reactions: knight427

Joules

EOS RP
Jul 16, 2017
323
251
Hamburg, Germany
I’d be shocked if any camera manufacturer ignores this reality and records data linearly.
Here's a quote from Adobe on this matter:

"Digital raw captures have linear gamma (gamma 1.0), a very different tonal
response from that of either film or the human eye. So the raw converter applies gamma correction to redistribute the tonal information so that it corresponds more closely to the way our eyes see light and shade"

RAW image Data is indeed linear data. It's called raw because the data is just what comes of the sensor: linear voltage, converted from photons. Twice as many photons = twice as bright in terms of what the camera sees.

Adobe source

We perceive doublings of brightness to be a significant change.
Is it not the opposite? The internet tells me that perceived brightness = sqrt(measured brightness) so a doubling in brightness (emitted energy) is not perceived as such by the human eye. To perceive a doubling of brightness, the emitted energy would have to be quadrupled.
 

dtaylor

Canon 5Ds
Jul 26, 2011
1,397
840
With a linear ADC, it's impossible to get more than 14 stops of DR from a 14-bit ADC.
The only way Sony can do it is to use a non-linear ADC.
Do any sensors exceed 14ev on DxO's screen test? The print test is normalized down, averaging out noise and boosting signal in 2 dimensions, which is how cameras get above 14. Is Sony doing the same in their published measurement?

Also: I could be dead wrong, but I was under the impression that it was very difficult to manufacture a linear ADC and that the ADCs etched into sensor silicon were most likely not linear.
 

knight427

EOS 80D
Aug 27, 2018
104
160
UOTE]

Is it not the opposite? The internet tells me that perceived brightness = sqrt(measured brightness) so a doubling in brightness (emitted energy) is not perceived as such by the human eye. To perceive a doubling of brightness, the emitted energy would have to be quadrupled.

My statement in no way conflicts with your sleuthing. I specifically did not quantify the "significance" of change, I was trying to keep the math simple and not get bogged down in unnecessary details. I was simply saying that a 2x change to linear brightness would be noticeable, whether that's a change of 1 to 2 or 5,000 to 10,000. I did not claim that such a change would be perceived as twice as bright.

I was honestly just guessing that a 2x linear change would be close to the threshold of perception for change in a non-laboratory observation. I based this on sound, for which 2x sound pressure (non-coherent) equates to +3 dB, which is the threshold of change for most humans to notice a change in loudness in the real world (1 dB is the threshold for lab tests).
 
Last edited:

Quarkcharmed

EOS 5DMkIV
Feb 14, 2018
519
370
Australia
www.michaelborisenko.com
It's been done before, and quite often in audio.

Remember Companders? Audio lines would have a frequency range of 300hz to 3000hz, so you would compress a wider range into that line, and at the other end you would expand it back to the original. It would give you a better frequency range, but at the expense of quality. This is a similar concept.
it's done at a cost of higher frequencies, same in images, compressing an extra stop of DR into higher values (top half) won't cause any significant quality loss.
 

Quarkcharmed

EOS 5DMkIV
Feb 14, 2018
519
370
Australia
www.michaelborisenko.com
Do any sensors exceed 14ev on DxO's screen test? The print test is normalized down, averaging out noise and boosting signal in 2 dimensions, which is how cameras get above 14. Is Sony doing the same in their published measurement?
That's why I take any DxO results with a grain of salt. photostophotos looks a bit more reliable, still their method shows the same Sony FF sensors in crop mode to lose ~1 stop of DR. I don't understand how simple cropping causes loss of DR. So this method seems to be ok only when used to compare sensors of the same size.
DxO used along with photostophotos gives cumulative impression.

Also: I could be dead wrong, but I was under the impression that it was very difficult to manufacture a linear ADC and that the ADCs etched into sensor silicon were most likely not linear.
On the contrary, afaik modern sensors mostly use linear ADCs. Linear is simpler to make.
 

Quarkcharmed

EOS 5DMkIV
Feb 14, 2018
519
370
Australia
www.michaelborisenko.com
Go open two RAW files at two different ISOs in ACR. Set to 0 Color NR the lower ISO one should just show color noise. The higher ISO one should have more color noise. Now apply the same amount of CNR to both. Tell me which one benefits more.
Such a test would not be applicable to what we were arguing about. We have to compare images of the same ISO, but different noise.
Those DR tests measure some 'acceptable' noise threshold. We shoot the very same scene with Canon and Sony sensors. All the same settings and ISO, except with Canon you shoot and underexpose by -3ev and with Sony by -3.8ev. Now in Lightroom with Canon, you push the shadows to +3ev, and with Sony, to +3.8ev, and still get the acceptable noise, and the tests claim the level of noise will be the same.

Now your claim is, that after doing all of the above (including the final pushing to +3 and +3.8ev), noise reduction will work better for Canon for some mysterious reason.
 

privatebydesign

Would you take advice from a cartoons stuffed toy?
Jan 29, 2011
7,907
1,050
119
Nevertheless that is how they arrive at the numbers for the two files in the dual RAW system. The main file is the sum of the two sub pixels, the secondary file is the reading from one set of sub pixels. That is why the benefit seen is not in the shadows, the traditional improvement of lowering the noise floor, but in the shadows as it is adding extra range of tonality at the top.

Of course this means the benefit is very scene and exposure specific, if the main RAW file has no blown pixels then there is zero advantage to having dual RAW files.
You wrote shadows there twice and I think you meant to say highlights? Having trouble deciphering your comment.
No I meant shadows, sorry if my explanation was confusing.

The normal way of increasing DR in a bit limited file is to lower the noise floor, that is you get more detail with less noise in the shadows. Canon don't take this route with the dual RAW output, they divide the highest exposure stop of DR into two and this allows an extra set of tonal values in the highlights.

If you want the maximum DR range from a normal single RAW system you expose such that you don't overexpose your highlights (or at the least minimize blown highlights to a non distracting level), this lets you dig around in the shadows for 'extra' DR, with the Canon dual RAW system you find that highlight exposure limit then over expose one stop, the excessive blown highlights in the main RAW file will be recoverable, and hold genuine detail, in the sub RAW file. You only get this benefit with the Canon system if you overexpose the main RAW, if you don't there is no additional information in the sub RAW.
 

Don Haines

Beware of cats with laser eyes!
Jun 4, 2012
8,149
1,681
Canada
it's done at a cost of higher frequencies, same in images, compressing an extra stop of DR into higher values (top half) won't cause any significant quality loss.
A lot depends on the individual image. For some it will make them better and for others it will make them worse. That said, it is a fairly easy software task to identify the images that do not need DR compression and then leave them alone. This gets you the best of both worlds.

It is an even easier task to just save them as 16 bit images and forget all about DR compression, which is easier, faster, and better.
 

Quarkcharmed

EOS 5DMkIV
Feb 14, 2018
519
370
Australia
www.michaelborisenko.com
A lot depends on the individual image. For some it will make them better and for others it will make them worse. That said, it is a fairly easy software task to identify the images that do not need DR compression and then leave them alone. This gets you the best of both worlds.

It is an even easier task to just save them as 16 bit images and forget all about DR compression, which is easier, faster, and better.
Yeah, but that Sony A7RIV compression is presumably done in ADC (re: non-linear ADC). And ADC is 14 bits. So it'll need to be decompressed after it becomes digital, maybe into 16 bits or 15 bits (as long as they claim 15-stop DR).
 

dtaylor

Canon 5Ds
Jul 26, 2011
1,397
840
Now your claim is, that after doing all of the above (including the final pushing to +3 and +3.8ev), noise reduction will work better for Canon for some mysterious reason.
That's not my claim, that's your strawman.

But by all means don't try it. Don't let observation of the real world inform or guide you in any way. Persist in your beliefs, whatever they may be, until the sun turns black and the Earth turns cold.
 

Quirkz

EOS 80D
Oct 30, 2014
175
103
Yeah, but that Sony A7RIV compression is presumably done in ADC (re: non-linear ADC). And ADC is 14 bits. So it'll need to be decompressed after it becomes digital, maybe into 16 bits or 15 bits (as long as they claim 15-stop DR).
If this is correct, it means they’re choosing data to throw away - so a lossy compression, and losing detail for some brightness values to gain a little more DR. Not sure I’d be happy with this trade off. It works quite well in audio compression since you can focus on the frequencies where the ear is most sensitive, and you’re not interested in the data outside that frequency. But for DR in images, as far as I understand it, is about retaining detail in deep shadows and highlights - not just a narrow band of mid tones.
 

Quarkcharmed

EOS 5DMkIV
Feb 14, 2018
519
370
Australia
www.michaelborisenko.com
If this is correct, it means they’re choosing data to throw away - so a lossy compression, and losing detail for some brightness values to gain a little more DR. Not sure I’d be happy with this trade off. It works quite well in audio compression since you can focus on the frequencies where the ear is most sensitive, and you’re not interested in the data outside that frequency. But for DR in images, as far as I understand it, is about retaining detail in deep shadows and highlights - not just a narrow band of mid tones.
It's known that the last brightest stop of an image is represented by 1/2 of all brightness levels available. So if we compress say 2 stops into the same amount of levels, the quality won't be worse than it is in midtones.

But I don't know whether Sony really uses a non-linear ADC. I'm just trying to figure out how they managed to squeeze 15 stops of DR into 14-bit ADC.
 

Don Haines

Beware of cats with laser eyes!
Jun 4, 2012
8,149
1,681
Canada
Yeah, but that Sony A7RIV compression is presumably done in ADC (re: non-linear ADC). And ADC is 14 bits. So it'll need to be decompressed after it becomes digital, maybe into 16 bits or 15 bits (as long as they claim 15-stop DR).
I suppose that one could design a non linear ADC, but most designs are ladder style. Going non linear would be a lot more complex and would (presumably) be slower and take more power to run.

Besides, sensors are linear, and the number of bits required to sample the results depends on the full well capacity of the sensor. To gain an extra bit, you need twice the well depth, and Sony (or Canon) don’t have that.