Is the E-M5 sensor response nonlinear?

[Anders W]

In a recent thread, Iliah Borg claimed that the response of the E-M5 sensor was nonlinear close to the clipping point. If this claim is correct, using ETTR might not yield optimal image quality. Instead, it might be preferable to stay some distance below the clipping point of the sensor.

Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded as has been the case on some recent Sony cameras.

http://forums.dpreview.com/forums/readflat.asp?forum=1041&message=41965851&changemode=1

After finally receiving my own E-M5 a couple of days ago, I decided to put these claims/suggestions to the test. I found an evenly illuminated white wall and shot it at various shutter speed from some way below the clipping point and up to the point where all channels had clipped completely. The lens used was the 45/1.8 stopped down to 5.6 and completely defocused so as to yield as tight (narrow) a light distribution as possible. I then used RawDigger to examine the average raw levels recorded.

The expected linear increase for each 1/3 EV change is 1.26 (the cubic root of 2). Below, I present the average levels for each channel from a point as close to saturation as possible and one EV down in steps of 1/3 EV along with the observed rate of change. As is readily seen, the response is in each case almost perfectly linear. Consequently, we can use ETTR on the E-M5 without fear of losing IQ due to nonlinearity.

Red channel
3338 1.26
2650 1.28
2068 1.26
1636

Green channel (average of G and G2)
3615 1.27 (about 40k pixels clipped)
2846 1.26
2258 1.27
1774

Blue
3599 1.28
2807 1.27
2217 1.27
1740

 

[Steen Bay]

In a recent thread, Iliah Borg claimed that the response of the E-M5 sensor was nonlinear close to the clipping point. If this claim is correct, using ETTR might not yield optimal image quality. Instead, it might be preferable to stay some distance below the clipping point of the sensor.

That would be kind of a waste, and fortunately your test results show that there's nothing to worry about. Another thing that would be nice to know is how much it's possible to expose beyond the clipping point, and 'recover' with a good result in LR4.

Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded as has been the case on some recent Sony cameras.

http://forums.dpreview.com/forums/readflat.asp?forum=1041&message=41965851&changemode=1

Interesting, since it would make it possible to have more than 12 stops of pixel-level ('screen') DR, even though the raw files are 'just' 12-bit. Don't know how to check that (nonlinear encoding).

After finally receiving my own E-M5 a couple of days ago, I decided to put these claims/suggestions to the test. I found an evenly illuminated white wall and shot it at various shutter speed from some way below the clipping point and up to the point where all channels had clipped completely. The lens used was the 45/1.8 stopped down to 5.6 and completely defocused so as to yield as tight (narrow) a light distribution as possible. I then used RawDigger to examine the average raw levels recorded.

The expected linear increase for each 1/3 EV change is 1.26 (the cubic root of 2). Below, I present the average levels for each channel from a point as close to saturation as possible and one EV down in steps of 1/3 EV along with the observed rate of change. As is readily seen, the response is in each case almost perfectly linear. Consequently, we can use ETTR on the E-M5 without fear of losing IQ due to nonlinearity.

Red channel
3338 1.26
2650 1.28
2068 1.26
1636

Green channel (average of G and G2)
3615 1.27 (about 40k pixels clipped)
2846 1.26
2258 1.27
1774

Blue
3599 1.28
2807 1.27
2217 1.27
1740

 

[Iliah Borg]

Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded as has been the case on some recent Sony cameras.

http://forums.dpreview.com/forums/readflat.asp?forum=1041&message=41965851&changemode=1

??

Red channel
3338 1.26

Green channel (average of G and G2)
3615 1.27 (about 40k pixels clipped)

Blue
3599 1.28

Why clipping points are so far apart? Why do you average channels, it is wrong. What was the amount of flare?

--
http://www.libraw.org/
http://www.RawDigger.com/

 

[Detail Man]

In a recent thread, Iliah Borg claimed that the response of the E-M5 sensor was nonlinear close to the clipping point. If this claim is correct, using ETTR might not yield optimal image quality. Instead, it might be preferable to stay some distance below the clipping point of the sensor.

Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded as has been the case on some recent Sony cameras.

http://forums.dpreview.com/forums/readflat.asp?forum=1041&message=41965851&changemode=1

After finally receiving my own E-M5 a couple of days ago, I decided to put these claims/suggestions to the test. I found an evenly illuminated white wall and shot it at various shutter speed from some way below the clipping point and up to the point where all channels had clipped completely. The lens used was the 45/1.8 stopped down to 5.6 and completely defocused so as to yield as tight (narrow) a light distribution as possible. I then used RawDigger to examine the average raw levels recorded.

What was the ISO Gain set to ? What was the "Max" (peak) saturation level in ADU's at that ISO ?

The expected linear increase for each 1/3 EV change is 1.26 (the cubic root of 2). Below, I present the average levels for each channel from a point as close to saturation as possible and one EV down in steps of 1/3 EV along with the observed rate of change. As is readily seen, the response is in each case almost perfectly linear. Consequently, we can use ETTR on the E-M5 without fear of losing IQ due to nonlinearity.

Red channel
3338 1.26
2650 1.28
2068 1.26
1636

Green channel (average of G and G2)
3615 1.27 (about 40k pixels clipped)
2846 1.26
2258 1.27
1774

Blue
3599 1.28
2807 1.27
2217 1.27
1740

Curious as to what the "Max" (peak) data was. Did it also conform as closely as the "Avg" data ?

 

[Iliah Borg]

What was the ISO Gain set to ? What was the "Max" (peak) saturation level in ADU's at that ISO ?

I would start with field uniformity; and I do not see why to use a lens at all.
Next, one needs to check the shutter speeds. It is not all that simple.

--
http://www.libraw.org/
http://www.RawDigger.com/

 

[bobn2]

Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded as has been the case on some recent Sony cameras.

I didn't suggest, I asked the question.

PS. The thread you linked maxed out so I couldn't follow up. This was left hanging:

Bob: If it is an Exmor then the DxO stauration level expressed as an ISO will be around 80.

Anders: The DxO saturation level is unlikely to be about 80. The data we have suggest it is about 130.

On course, the saturation level that you see depends on the the raw file full scale, not the sensor's. What is unknown is Olympus use of the programmable variable gain amplifier in each columns (which does not necessarily go directly with ISO) and, if they use the 14 bit ADC, how they are mapping that to a 12 bit raw file, whether it is simple truncation (top, and/or bottom) or something more subtle. Sometimes some of that can be deduced, as Marianne Oelund has done for the Nikon cameras, which show a very precise tweaking of the gain and consequently different digital mappings from ISO to ISO.
--
Bob

 

[bobn2]

Interesting, since it would make it possible to have more than 12 stops of pixel-level ('screen') DR, even though the raw files are 'just' 12-bit. Don't know how to check that (nonlinear encoding).

You can in any case. 12 bits can encode more than 12 stops of DR, it just can't do it completely. Remember noise is a statistical variation, it is not something that is measured in a single pixel. So, it's not the case that, say 0.25 bits of noise will be registered as 1 bit. If you average a lot of bits in which there is 0.25 bits of noise, you'll find that the average value is 0.25, composed of three times as many '0' bits as '1' bits, though obviously randomly arranged. However, you can't encode fully the brightness ranges that should be available between '1' and '0.25' in a pixel. You can in large areas, where the noise dithers the average to give a smooth(ish) tone showing less than 1 bit's worth of brightness.
--
Bob

 

[Iliah Borg]

On course, the saturation level that you see depends on the the raw file full scale,

Which, for the cameras I profiled, was 3791 (black level subtracted) uniformly between all 4 channels.

What is unknown is Olympus use of the programmable variable gain amplifier in each columns (which does not necessarily go directly with ISO) and, if they use the 14 bit ADC, how they are mapping that to a 12 bit raw file, whether it is simple truncation (top, and/or bottom) or something more subtle.

Sometimes it is good to check that SNR at exp1 to SNR at exp2 is along with the linear model.

On a side note, Nikon D4 has a strong non-linearity at the highlight portion of the characteristic curve which is compensated by calibration of sensor assemblies. When this calibration is not done properly the camera is not very easy to use at daylight - one needs to underexpose 2 stops to get decent output.

--
http://www.libraw.org/
http://www.RawDigger.com/

 

[Steen Bay]

What was the ISO Gain set to ? What was the "Max" (peak) saturation level in ADU's at that ISO ?

Guess it was at base ISO, otherwise the test wouldn't tell much.

I would start with field uniformity; and I do not see why to use a lens at all.
Next, one needs to check the shutter speeds. It is not all that simple.

It probably ain't that simple, but on the other hand, isn't it almost impossible to get such an almost perfect match by pure coincidence?

 

[Steen Bay]

Interesting, since it would make it possible to have more than 12 stops of pixel-level ('screen') DR, even though the raw files are 'just' 12-bit. Don't know how to check that (nonlinear encoding).

You can in any case. 12 bits can encode more than 12 stops of DR, it just can't do it completely. Remember noise is a statistical variation, it is not something that is measured in a single pixel. So, it's not the case that, say 0.25 bits of noise will be registered as 1 bit. If you average a lot of bits in which there is 0.25 bits of noise, you'll find that the average value is 0.25, composed of three times as many '0' bits as '1' bits, though obviously randomly arranged. However, you can't encode fully the brightness ranges that should be available between '1' and '0.25' in a pixel. You can in large areas, where the noise dithers the average to give a smooth(ish) tone showing less than 1 bit's worth of brightness.

Isn't that the same as saying that DR increases when downsampling an image? Anyway, wouldn't it be a good idea to use nonlinear encoding, even if a camera with 12-bit raw files just has for example 10.7 stops of DR? Isn't it a big waste to use half of the available levels for the first stop?

 

[Iliah Borg]

It probably ain't that simple, but on the other hand, isn't it almost impossible to get such an almost perfect match by pure coincidence?

The max values are off compared to several camera samples I profiled. So I do not see a match. Noise analysis is absent, RawDigger mode is not stated...

Finally, in the thread Anders referrs to I wrote:
"The output to raw from E-M5 is rather linear."

http://forums.dpreview.com/forums/read.asp?forum=1041&message=41971826 - which does not exactly the same as his: "Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded".

--
http://www.libraw.org/
http://www.RawDigger.com/

 

[Anders W]

Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded as has been the case on some recent Sony cameras.

http://forums.dpreview.com/forums/readflat.asp?forum=1041&message=41965851&changemode=1

??

In case you have already forgotten our discussion in the other thread, here's the beginning of it:

Iliah Borg: The "brightest" stop is plagued with noise, so for such a scene you may want to keep in "ETTR-1".

Anders W: On what grounds do you consider the "brightest" stop plagued with noise? Because there is a partial risk of clipping? If so, then you aren't really in the brightest stop but a bit beyond.

Iliah Borg: Sensors are not linear close to saturation, shot noise is dominating in the lighter parts of the image; modern raw compression schemes use much less levels to represent lighter parts of the image because the information there is already unreliable.

Red channel
3338 1.26

Green channel (average of G and G2)
3615 1.27 (about 40k pixels clipped)

Blue
3599 1.28

Why clipping points are so far apart?

The clipping point of the sensor is the same in each channel (3815). These are the averages I recorded when exposing as close to that point as possible in 1/3 EV increments (rather than in a perfectly continuous manner). Unless you have a perfectly even mix of red, green, and blue light at your disposal, exactly how close you get will vary slightly from one channel to the other.

Why do you average channels, it is wrong.

1. Why is it wrong to average the two green channels (which are the only ones I averaged)?

2. The results when not averaging these two channels are of course substantively identical to the averaged result.

What was the amount of flare?

What kind of flare are you talking about and why would that matter?

 

[Steen Bay]

It probably ain't that simple, but on the other hand, isn't it almost impossible to get such an almost perfect match by pure coincidence?

The max values are off compared to several camera samples I profiled. So I do not see a match. Noise analysis is absent, RawDigger mode is not stated...

Don't think it's supposed to be the sensors max values, just as close as it was possible to get (without clipping) when changing the shutterspeed in 1/3 Ev steps.

Finally, in the thread Anders referrs to I wrote:

"The output to raw from E-M5 is rather linear."

http://forums.dpreview.com/forums/read.asp?forum=1041&message=41971826 - which does not exactly the same as his: "Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded".

--
http://www.libraw.org/
http://www.RawDigger.com/

 

[Detail Man]

The max values are off compared to several camera samples I profiled. So I do not see a match.

Gakuranman 's EM5 (overexposed, saturated in all RAW channels) at ISO=200 maxed at 3811.

texinwien (similarly overexposed, saturated in all RAW channels) measured the following (RGBG):

200 ------- 3815 3815 3815 3816
400 ------- 3841 3841 3841 3841
800 ------- 3841 3842 3841 3842
1600 ----- 3844 3845 3844 3846
3200 ----- 3852 3853 3852 3854
6400 ----- 3867 3870 3868 3872
12800 --- 3894 3902 3897 3905
25600 --- 3947 3961 3953 3967

However, these were all "Max" (peak level) statistics - whereas AndersW reported "Ave" statistics. The ISO Gain that the data was collected at is not stated. "Max" stats appear to increase with ISO

 

[Anders W]

In a recent thread, Iliah Borg claimed that the response of the E-M5 sensor was nonlinear close to the clipping point. If this claim is correct, using ETTR might not yield optimal image quality. Instead, it might be preferable to stay some distance below the clipping point of the sensor.

Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded as has been the case on some recent Sony cameras.

http://forums.dpreview.com/forums/readflat.asp?forum=1041&message=41965851&changemode=1

After finally receiving my own E-M5 a couple of days ago, I decided to put these claims/suggestions to the test. I found an evenly illuminated white wall and shot it at various shutter speed from some way below the clipping point and up to the point where all channels had clipped completely. The lens used was the 45/1.8 stopped down to 5.6 and completely defocused so as to yield as tight (narrow) a light distribution as possible. I then used RawDigger to examine the average raw levels recorded.

What was the ISO Gain set to ?

ISO 200 of course.

What was the "Max" (peak) saturation level in ADU's at that ISO ?

3815 is the clipping point.

The expected linear increase for each 1/3 EV change is 1.26 (the cubic root of 2). Below, I present the average levels for each channel from a point as close to saturation as possible and one EV down in steps of 1/3 EV along with the observed rate of change. As is readily seen, the response is in each case almost perfectly linear. Consequently, we can use ETTR on the E-M5 without fear of losing IQ due to nonlinearity.

Red channel
3338 1.26
2650 1.28
2068 1.26
1636

Green channel (average of G and G2)
3615 1.27 (about 40k pixels clipped)
2846 1.26
2258 1.27
1774

Blue
3599 1.28
2807 1.27
2217 1.27
1740

Curious as to what the "Max" (peak) data was.

See above.

Did it also conform as closely as the "Avg" data ?

I didn't bother to additionally compare peak to peak since the averages (with less measurement error) were so close to the peak. That was the point of setting up the little experiment in the manner I did.

 

[Anders W]

It probably ain't that simple, but on the other hand, isn't it almost impossible to get such an almost perfect match by pure coincidence?

The max values are off compared to several camera samples I profiled. So I do not see a match. Noise analysis is absent, RawDigger mode is not stated...

Don't think it's supposed to be the sensors max values, just as close as it was possible to get (without clipping) when changing the shutterspeed in 1/3 Ev steps.

Exactly!

Finally, in the thread Anders referrs to I wrote:

"The output to raw from E-M5 is rather linear."

http://forums.dpreview.com/forums/read.asp?forum=1041&message=41971826 - which does not exactly the same as his: "Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded".

--
http://www.libraw.org/
http://www.RawDigger.com/

 

[Anders W]

What was the ISO Gain set to ? What was the "Max" (peak) saturation level in ADU's at that ISO ?

I would start with field uniformity;

Why?

and I do not see why to use a lens at all.

Why not?

Next, one needs to check the shutter speeds.

Why?

It is not all that simple.

If it isn't and you can prove me wrong, please show the better data and the results that you claim you possess but doesn't show in spite of repeated specific questions about the matter in the other thread where the discussion started out.

 

[Iliah Borg]

Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded as has been the case on some recent Sony cameras.

http://forums.dpreview.com/forums/readflat.asp?forum=1041&message=41965851&changemode=1

??

In case you have already forgotten our discussion in the other thread, here's the beginning of it:

Iliah Borg: The "brightest" stop is plagued with noise, so for such a scene you may want to keep in "ETTR-1".

Anders W: On what grounds do you consider the "brightest" stop plagued with noise? Because there is a partial risk of clipping? If so, then you aren't really in the brightest stop but a bit beyond.

Iliah Borg: Sensors are not linear close to saturation, shot noise is dominating in the lighter parts of the image; modern raw compression schemes use much less levels to represent lighter parts of the image because the information there is already unreliable.

Best case, you do not see the difference between nonlinearly encoded and not linear by nature. That happens....

--
http://www.libraw.org/
http://www.RawDigger.com/

 

[Anders W]

Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded as has been the case on some recent Sony cameras.

I didn't suggest, I asked the question.

You said we couldn't exclude either alternative. So I tried to check whether we could.

PS. The thread you linked maxed out so I couldn't follow up. This was left hanging:

Bob: If it is an Exmor then the DxO stauration level expressed as an ISO will be around 80.

Anders: The DxO saturation level is unlikely to be about 80. The data we have suggest it is about 130.

On course, the saturation level that you see depends on the the raw file full scale, not the sensor's. What is unknown is Olympus use of the programmable variable gain amplifier in each columns (which does not necessarily go directly with ISO) and, if they use the 14 bit ADC, how they are mapping that to a 12 bit raw file, whether it is simple truncation (top, and/or bottom) or something more subtle. Sometimes some of that can be deduced, as Marianne Oelund has done for the Nikon cameras, which show a very precise tweaking of the gain and consequently different digital mappings from ISO to ISO.

Thanks. Yes, I realize this might be the explanation. Still it puzzles me why Oly would consciously abstain from using as low an ISO as the sensor is capable of (as does Nikon, Pentax, and Sony).

 

[Anders W]

Both Iliah Borg and bobn2 additionally suggested that the sensor response might be nonlinearly encoded as has been the case on some recent Sony cameras.

http://forums.dpreview.com/forums/readflat.asp?forum=1041&message=41965851&changemode=1

??

In case you have already forgotten our discussion in the other thread, here's the beginning of it:

Iliah Borg: The "brightest" stop is plagued with noise, so for such a scene you may want to keep in "ETTR-1".

Anders W: On what grounds do you consider the "brightest" stop plagued with noise? Because there is a partial risk of clipping? If so, then you aren't really in the brightest stop but a bit beyond.

Iliah Borg: Sensors are not linear close to saturation, shot noise is dominating in the lighter parts of the image; modern raw compression schemes use much less levels to represent lighter parts of the image because the information there is already unreliable.

Best case, you do not see the difference between nonlinearly encoded and not linear by nature. That happens....

Sure. And so what?