First impressions - A7R III preproduction unit

[bclaff]

...
... Bill Claff has chosen a SNR of 20%, ... Sometimes a SNR of 20 ... his assumption is that 14 stops of DR is a limit in 14 bit files and his calculations are based on 14 stops minus the amount lost to the 20% SNR....

It's an SNR of 20; not an SNR of 20%.
There is no assumption that 14-bit ADC limits dynamic range to 14.

 

[bclaff]

...

FWIW.
The Sony claim cites neither CIPA nor DxOMark; it has no concrete foundation.
It simply says "Sony test conditions for stills"; whatever that means.

 

[Jopa]

...

FWIW.
The Sony claim cites neither CIPA nor DxOMark; it has no concrete foundation.
It simply says "Sony test conditions for stills"; whatever that means.

That's funny. Next time they'll simply say "because we say so" ;D

 

[privatebydesign]

There is no assumption that 14-bit ADC limits dynamic range to 14.

Then I don't understand why your explanation page runs to 14 and your figures are 14 minus the amount lost to < SNR 20.
"Here is a chart of the Photon Transfer Curve for the Nikon D300 in ADUs at all whole ISOs based on 14-bit ADUs: "

http://www.photonstophotos.net/GeneralTopics/Sensors_&_Raw/Sensor_Analysis_Primer/Engineering_and_Photographic_Dynamic_Range.htm

I understand 14 bit ADU's can represent any range of values with 16,384 steps, and that an ADC doesn't count 1:1 photons to these steps, what I don't understand is where the 14 stops of DR as a base comes from.

So, for example, a full well capacity for a pixel is 100,000 photons. The 14 bit ADC can change that 0-100,000 photons of electric charge into any one of 16,384 values, or ADU's, giving a 14 bit ADU. I understand that if it took 20 stops of DR to capture a full range of photons across the sensor then the 16,000 steps would represent 20 stops of DR, but that doesn't seem to be how it works and again I don't understand where your 14 stop start point comes from.

Would be grateful of an explanation from somebody who actually understands!

 

[bwud]

...

FWIW.
The Sony claim cites neither CIPA nor DxOMark; it has no concrete foundation.
It simply says "Sony test conditions for stills"; whatever that means.

I know, I’m basing my premise on their CIPA membership.

 

[bclaff]

There is no assumption that 14-bit ADC limits dynamic range to 14.

Then I don't understand why your explanation page runs to 14 and your figures are 14 minus the amount lost to < SNR 20.
"Here is a chart of the Photon Transfer Curve for the Nikon D300 in ADUs at all whole ISOs based on 14-bit ADUs: "

http://www.photonstophotos.net/GeneralTopics/Sensors_&_Raw/Sensor_Analysis_Primer/Engineering_and_Photographic_Dynamic_Range.htm

I understand 14 bit ADU's can represent any range of values with 16,384 steps, and that an ADC doesn't count 1:1 photons to these steps, what I don't understand is where the 14 stops of DR as a base comes from.

So, for example, a full well capacity for a pixel is 100,000 photons. The 14 bit ADC can change that 0-100,000 photons of electric charge into any one of 16,384 values, or ADU's, giving a 14 bit ADU. I understand that if it took 20 stops of DR to capture a full range of photons across the sensor then the 16,000 steps would represent 20 stops of DR, but that doesn't seem to be how it works and again I don't understand where your 14 stop start point comes from.

Would be grateful of an explanation from somebody who actually understands!

This comes up enough that a separate article on PhotonsToPhotos is called for.

First note that the chart you make reference to has a logarithmic x-axis.
It's true it ends on the right at 14 but it also runs to minus infinity on the left even though that particular one ends at 0.

Remember, read noise is a standard deviation; so although each ADC reading is integral the standard deviation is not, and can be less than 1.
Because of quantization error in practice we can't measure below about 0.6 DN
( see http://www.photonstophotos.net/GeneralTopics/Sensors_&_Raw/Quantization_Error_in_Practice.htm )
0.6 is nearly 0.75 stops better than you might expect but to be conservative I typically say that an n-bit ADC can measure n+0.5 stops of dynamic range; 14.5 stops for a 14-bit.

I rarely test a sensor with 14-bit ADC with more than 14 stops of EDR; the Nikon D7200 is an example.
This happens more often with 12-bit ADCs; I assume these cameras don't use 14-bit due to cost considerations.
A 12-bit example is the Panasonic GX80.
Even the 10-bit ADC in the Samsung S6 phone isn't enough at ISO 50.
Attaching (note logarithmic y-axis, below 0 is below 1DN ):
http://www.photonstophotos.net/Charts/RN_ADU.htm#Nikon%20D7200_14,Panasonic%20Lumix%20DMC-GX80_12,Samsung%20Galaxy%20S6(S5K2P2)_10

 

[Mikehit]

[
This comes up enough that a separate article on PhotonsToPhotos is called for.

First note that the chart you make reference to has a logarithmic x-axis.
It's true it ends on the right at 14 but it also runs to minus infinity on the left even though that particular one ends at 0.

Remember, read noise is a standard deviation; so although each ADC reading is integral the standard deviation is not, and can be less than 1.
Because of quantization error in practice we can't measure below about 0.6 DN
( see http://www.photonstophotos.net/GeneralTopics/Sensors_&_Raw/Quantization_Error_in_Practice.htm )
0.6 is nearly 0.75 stops better than you might expect but to be conservative I typically say that an n-bit ADC can measure n+0.5 stops of dynamic range; 14.5 stops for a 14-bit.

I rarely test a sensor with 14-bit ADC with more than 14 stops of EDR; the Nikon D7200 is an example.
This happens more often with 12-bit ADCs; I assume these cameras don't use 14-bit due to cost considerations.
A 12-bit example is the Panasonic GX80.
Even the 10-bit ADC in the Samsung S6 phone isn't enough at ISO 50.
Attaching (note logarithmic y-axis, below 0 is below 1DN ):
http://www.photonstophotos.net/Charts/RN_ADU.htm#Nikon%20D7200_14,Panasonic%20Lumix%20DMC-GX80_12,Samsung%20Galaxy%20S6(S5K2P2)_10

I don't know enough about the technical stuff going on here, but on first reading it is starting to sound like there is so much electronic jiggery-pokery along with each company's interpretation of the definitions that it is becoming hard to believe what any company is saying any more because there is no hard and fast definition of what constitutes acceptable noise, or dynamic range.
The MP wars have got to the point that 30MP vs 45 MP? Who really cares. And dynamic range seems pretty much the same. Everyone is fighting over scraps at the table

 

[privatebydesign]

..... there is no hard and fast definition of what constitutes acceptable noise, or dynamic range.
The MP wars have got to the point that 30MP vs 45 MP? Who really cares. And dynamic range seems pretty much the same. Everyone is fighting over scraps at the table

Yep, and I keep pointing that out, not that the forum DRones or measurbators care........

http://www.canonrumors.com/forum/index.php?topic=33696.msg694260#msg694260

http://www.canonrumors.com/forum/index.php?topic=33709.msg693866#msg693866

 

[bwud]

... because there is no hard and fast definition of what constitutes acceptable noise, or dynamic range.

Oh yes there is. The trade group the major camera makers (including Canon, Fuji, Nikon, Sony, and Ricoh) belong to has a (non-mandatory) guideline as to how camera specifications should be evaluated and expressed in order to "...inform consumers of product specifications fairly and accurately and to develop a sound market and digital camera business by encouraging fair competition..." That guideline includes dynamic range, in specific:

Shall comply with ISO 15739:2003 (an outline is shown below).
1. Definition
This means the ratio of the maximum level of luminance signals without saturation to the level of luminance signals where the S/N ratio to random noise is 1 (the minimum level of luminance signals)

Many people don't like that definition applied to camera products, and with good reason, but rather than petitioning ISO to change the standard, they come up the myriad hodgepodge of formulae we see in the marketplace (mostly in testing websites since few vendors publish DR).

However, there most certainly and definitively exists a hard and fast definition of what constitutes dynamic range: ISO 15739:2003 Photography -- Electronic still-picture imaging -- Noise measurements. Additionally, there is commercially available software to test per ISO, such as Imatest.

I pose a question to Bill Claff: if you defined dynamic range per the ISO standard, with the data you collected, what value would you compute?

 

[privatebydesign]

Many people don't like that definition applied to camera products, and with good reason, but rather than petitioning ISO to change the standard, they come up the myriad hodgepodge of formulae we see in the marketplace (mostly in testing websites since few vendors publish DR).

When Canon did that with their Cinema cameras they were absolutely hammered over it.

As for the Claff ISO standard, surely in his work that equates to the Engineering Dynamic Range.

 

[bwud]

Many people don't like that definition applied to camera products, and with good reason, but rather than petitioning ISO to change the standard, they come up the myriad hodgepodge of formulae we see in the marketplace (mostly in testing websites since few vendors publish DR).

When Canon did that with their Cinema cameras they were absolutely hammered over it.

As for the Claff ISO standard, surely in his work that equates to the Engineering Dynamic Range.

You’re probably right (his EDR figure). I’m just curious if Sony has any leg to stand on, assuming its data isn’t dramatically different, WRT the claimed 15-stops (without relying on noise reducing downsampling). My guess is no, but then again as you note Canon published data, and though they took heat for it, they backed it up with an entirely different method (instrumentation).

 

[bclaff]

... because there is no hard and fast definition of what constitutes acceptable noise, or dynamic range.

Oh yes there is. The trade group the major camera makers (including Canon, Fuji, Nikon, Sony, and Ricoh) belong to has a (non-mandatory) guideline as to how camera specifications should be evaluated and expressed in order to "...inform consumers of product specifications fairly and accurately and to develop a sound market and digital camera business by encouraging fair competition..." That guideline includes dynamic range, in specific:

Shall comply with ISO 15739:2003 (an outline is shown below).
1. Definition
This means the ratio of the maximum level of luminance signals without saturation to the level of luminance signals where the S/N ratio to random noise is 1 (the minimum level of luminance signals)

...
I pose a question to Bill Claff: if you defined dynamic range per the ISO standard, with the data you collected, what value would you compute?

There no need to cite any ISO or other standard; dynamic range at the pixel level is always defined that way.
Because there are other normalized dynamic range measures I always call this Engineering Dynamic Range (EDR).
For the 7R Mark III I have complete data; attached below (and compared with the Mark II)
( http://www.photonstophotos.net/Charts/RN_ADU.htm#Sony%20ILCE-7RM3_14 )
The numbers are shown (mouse overe data points) with too much precision but the answer is unaffected.
log2((16383DN-512DN)/1.257DN) = 13.6 stops
512DN is known as BlackLevel.
1.257DN is the lowest read noise (measured at ISO 50); but even if you substitute 1.275DN at ISO 100 you still get 13.6 stops.
Note this isn't close to 15, nor is any existing camera; hence the assumption that this can't be what Sony means.
BTW, for the 7R Mark II it comes out 13.3 stops; so the Mark III is a 0.3 stop improvement in EDR.

 

[bwud]

Thanks Bill.

 

[Talys]

Thanks Bill.

+1 ... interesting, and thanks, Bill!

 

[snoke]

There is no assumption that 14-bit ADC limits dynamic range to 14.

Then I don't understand why your explanation page runs to 14 and your figures are 14 minus the amount lost to < SNR 20.
"Here is a chart of the Photon Transfer Curve for the Nikon D300 in ADUs at all whole ISOs based on 14-bit ADUs: "

http://www.photonstophotos.net/GeneralTopics/Sensors_&_Raw/Sensor_Analysis_Primer/Engineering_and_Photographic_Dynamic_Range.htm

I understand 14 bit ADU's can represent any range of values with 16,384 steps, and that an ADC doesn't count 1:1 photons to these steps, what I don't understand is where the 14 stops of DR as a base comes from.

So, for example, a full well capacity for a pixel is 100,000 photons. The 14 bit ADC can change that 0-100,000 photons of electric charge into any one of 16,384 values, or ADU's, giving a 14 bit ADU. I understand that if it took 20 stops of DR to capture a full range of photons across the sensor then the 16,000 steps would represent 20 stops of DR, but that doesn't seem to be how it works and again I don't understand where your 14 stop start point comes from.

Would be grateful of an explanation from somebody who actually understands!

This comes up enough that a separate article on PhotonsToPhotos is called for.

First note that the chart you make reference to has a logarithmic x-axis.
It's true it ends on the right at 14 but it also runs to minus infinity on the left even though that particular one ends at 0.

Remember, read noise is a standard deviation; so although each ADC reading is integral the standard deviation is not, and can be less than 1.
Because of quantization error in practice we can't measure below about 0.6 DN
( see http://www.photonstophotos.net/GeneralTopics/Sensors_&_Raw/Quantization_Error_in_Practice.htm )
0.6 is nearly 0.75 stops better than you might expect but to be conservative I typically say that an n-bit ADC can measure n+0.5 stops of dynamic range; 14.5 stops for a 14-bit.

I rarely test a sensor with 14-bit ADC with more than 14 stops of EDR; the Nikon D7200 is an example.
This happens more often with 12-bit ADCs; I assume these cameras don't use 14-bit due to cost considerations.
A 12-bit example is the Panasonic GX80.
Even the 10-bit ADC in the Samsung S6 phone isn't enough at ISO 50.
Attaching (note logarithmic y-axis, below 0 is below 1DN ):

Who make Panasonic sensor? GX80 read noise under FF DSLR, yes?

 

[bclaff]

...
GX80 read noise under FF DSLR, yes?

Comparing per pixel read noise in DN is senseless.

 

[snoke]

...
GX80 read noise under FF DSLR, yes?

Comparing per pixel read noise in DN is senseless.

Don't understand presentation if senseless. How to use graph?

 

[tron]

hello,

the comments below are addressed to Sony and not yourself of course.

On the contrary I would like to thank you for providing us with your detailed feedback.

...
In addition, for some reason, the viewfinder started to stutter in jarring (unusable) way when the camera was in continuous autofocus. When I pointed it out to the Sony guy, he reset the camera, and that went away. He blamed it on preproduction firmware, which is entirely plausible....

It is plausible but not certain.

He also said that they had a number of other issues with that unit.

Alarming...

 

[tron]

Final thoughts -- well, at this moment, even if I really wanted one, after seeing the preproduction unit, I sure wouldn't buy one, until it came out and I was sure the bugs were all ironed out. Boy, were there a lot, for a unit that they were demonstrating to the general public. Maybe I'm being unfair, but it feels like it's rushed to make Christmas.

+ 1000