304 replies to this topic

[sharkmelley]

I am starting a new thread for my Nikon Z6 testing. 

 

It will take me a few weeks to run all the tests I need to perform and to write a proper review but I'll post some early findings here.

 

Meanwhile anyone interested in the Z6 should check out Alan Dyer's review because I don't intend to duplicate what he has already done:
https://amazingsky.n...trophotography/

 

 

Read Noise, Gain etc.

 

I also don't intend to reproduce Bill Claff's excellent work:

• Read Noise:    http://www.photonsto...tm#Nikon Z 6_14
• Dynamic Range:   http://www.photonsto...R.htm#Nikon Z 6

 

I have arbitrarily decided to perform most testing at ISO 800, which is where the sensor switches to HGC (high gain) mode.  At ISO 800, I determined the read noise to be somewhere in the range 1.67 to 1.93e RMS.  Bill reports  1.705e.  My uncertainty is caused by the fact that successive bias frames often have histograms with gaps in different places.  This odd behaviour affects the measured standard deviation and hence the implied read noise.

 

 

Quantum Efficiency (QE)

 

One of the things that really interested me about the Nikon Z6 was the quantum efficiency.  Some time ago I developed a way of analysing the raw files that can be downloaded from DPReview's Studio scene to deduce camera gain (i.e. the conversion factor from pixel values to electrons) and hence the number of captured photons.

e.g. https://www.dpreview...656129343629345

 

To compare cameras using this analysis relies on several assumptions:

• the scene is always illuminated with identical brightness
• the focal ratio of the lens being used is accurate
• the reported shutter speed of the camera is accurate 

The results suggested that the back side illuminated (BSI) sensor in the Nikon Z6 was capturing 14% more photons in the green channel than the existing low-light king, the Sony A7S.  Single value QE figures typically use the green channel.

 

I have now tested the 2 cameras with the same fully manual lens i.e. the lens has a good old aperture ring. I get the same result - the Z6 is capturing 14% more green photons that the A7S.  However, I am not an expert in sensor construction so I can't tell if 14% is a reasonable increase for BSI technology.

 

If it is correct that the Z6 has a higher QE then it means that the Nikon Z6 should outperform the renowned Sony A7S in situations where sky fog is the dominant source of noise i.e. typical long exposure astro-imaging.

However, for situations dominated by read noise (e.g. low-light video) the Sony A7S will be impossible to beat.  The slowest frame rate on the Z6 is 25 frames/sec compared with 4 frames/sec on the A7S. So I can't do the trick of composing the scene by briefly switching to Movie mode at high ISO and low frame rate.

 

Mark

Edited by sharkmelley, 28 July 2019 - 06:38 AM.

[sharkmelley]

Dark Current and Sensor Self-Heating

 

The Nikon Z6 has a sensor with IBIS (in body image stabilization).  This means the sensor must be free to be moved, which might make it more difficult to conduct away heat.  To investigate, I took my usual approach of placing the camera in a dark room at 20C ambient temperature and taking successive 5 minute dark frames on a camera that has cooled to ambient.  The dark current is estimated by subtracting each one from its successor (which removes most of the fixed pattern noise).

 

For this experiment IBIS was switched off and so was the rear display.  Most cameras feel pretty warm at the back after 2 hours of successive exposures but this was not the case with the Nikon Z6.  The graphs bear this out.

 

Firstly here is the graph of per/pixel dark current, which is useful for comparing dark current with other sources of pixel noise e.g. read noise:

 

 

 

The second graph shows the dark current per square micron of sensor:

 

 

I find the second graph more useful because it makes it easier to compare between sensors, all having different pixel sizes.  Despite having IBIS, the sensor self-heating is lower than almost every camera I have tested - the Sony A7S being the remarkable exception.

 

One caveat I should add is that I have not yet adjusted the Nikon Z6 dark current figures to take account of the effects of spatial filtering on image noise.  This will require a very in-depth an analysis of the spatial filtering, which will come later.  The final dark current figures for the Nikon Z6 could be up to 20% higher. 

 

Mark

Edited by sharkmelley, 28 July 2019 - 05:41 AM.

[sharkmelley]

I should point out that I am doing most of my testing using PixInsight.  PixInsight does not yet support raw Nikon Z6 NEF files directly because they are not yet supported in a release version of LibRaw.  So I use Adobe DNG Converter to create DNG versions of all my NEF raw files.  This is a surprisingly fast operation and PixInsight supports DNG.

 

PDAF Banding

 

There are many online discussions about potential stripes in NIkon Z6 images caused by the on-sensor phase detect autofocus (PDAF) pixels.  The present consensus on DPReview's Nikon Z Mirrorless forum is that the stripes are actually caused by Nikon using a firmware algorithm to avoid the PDAF banding seen on other mirrorless cameras e.g. Sony.

 

Whatever the truth of the cause, my interest is whether or not this might affect astro-images, especially images processed using (bias/dark/flat) calibrated exposures.  I won't have a proper answer to this until I get a clear sky to use the camera on a telescope.  But here is what I have discovered so far.

 

PDAF banding is very obvious in a stretched master bias created from 25 exposures:

 

 

 

It is hidden by the noise in a stretched master dark created from 25 exposures of 5 minutes each:

 

 

 

If I run the PixInsight superbias process on the master dark it emphasises the horizontal structures and makes the banding visible:

 

 

 

So the banding exists in both bias and darks. The good news is that when the master bias is subtracted from the master dark then the PDAF striping is undetectable even after executing superbias on the resulting image i.e. the stripes appear to calibrate out.  This is certainly encouraging but the final verdict must await testing under real imaging conditions.

 

Mark

Edited by sharkmelley, 28 July 2019 - 06:30 AM.

[whwang]

Hi Mark,

 

It looks good so far.

 

One interesting thing is whether the PDAF pixels have higher or lower read noise than the others.  Do you think your tests can reveal that?

 

Cheers,

Wei-Hao

[sharkmelley]

Hi Mark,

 

It looks good so far.

 

One interesting thing is whether the PDAF pixels have higher or lower read noise than the others.  Do you think your tests can reveal that?

 

Cheers,

Wei-Hao

Yes it could be done.

 

Bill Claff has performed such an analysis for the Nikon Z7:

https://www.dpreview...s/post/61542680

 

He concludes that the Z7 PDAF pixels are slightly noisier than the non-PDAF pixels.   My guess is that the results for the Z6 would be similar.

 

Mark

[whwang]

Thanks.  Looks like they are about 10% noisier, which is not something to worry about.

 

Next would be to see if they are hotter.

[Jerry Lodriguss]

 

Quantum Efficiency (QE)

The results suggested that the back side illuminated (BSI) sensor in the Nikon Z6 was capturing 14% more photons in the green channel than the existing low-light king, the Sony A7S. 

Hi Mark,

 

I'm looking at Bill Claff's data, and he has the Z6 at 57% and the A7S (assuming the ILCE-7S is the A7S) is at 62%.

 

Roger Clark seems to think that difference is trivial.

 

I think every little bit helps.

 

You say the A7S was the low light king, in a sentence referencing QE. But if you sort Bill's data by QE (just click on the "QE" at the top of the column to sort ascending or descending), there are 50 cameras with a higher listed QE.

 

I don't know how to interpret this in light of your statement.

 

I'm not challenging it, I just don't understand the QE numbers on Bill's site.

 

Is this full-system camera QE we are talking about including microlenses, BSI, etc?

 

Is this QE calculated for the full frame, or just on a pixel basis?

 

Is the data normalized for pixel size if it's on a per pixel basis?

 

In general conceptual terms, how is QE calculated from DxO data?

 

I ask, because some there are 6 cameras listed with 100% or higher QE.

 

I've always been skeptical of DxO data because of discrepancies like this. Maybe there are reasons for this that explain it, but I thought I went looking for an explanation years ago and everything was opaque at DxO.  Or maybe I'm just not smart enough to understand it.

 

Jerry 

Edited by Jerry Lodriguss, 28 July 2019 - 11:28 AM.

[sharkmelley]

Hi Jerry,

 

Lots of questions!  I'll try to answer as best I can:

 

 

I'm looking at Bill Claff's data, and he has the Z6 at 57% and the A7S (assuming the ILCE-7S is the A7S) is at 62%. 

Those values are derived from DXO data.  Bill has posted on DPReview some updated values where both A7S and Z6 are 58%:

https://www.dpreview...s/post/61988234

I don't understand the methodology involved in producing absolute QE levels so I concentrate on producing relative levels.

 

 

Roger Clark seems to think that difference is trivial.

I think every little bit helps.

Every little bit certainly helps but even a relative difference of 14% will be hardly noticeable in a final image.  If my figure of 14% is wrong then the Z6 is almost certainly as good as the Sony A7S.

 

 

You say the A7S was the low light king, in a sentence referencing QE. But if you sort Bill's data by QE (just click on the "QE" at the top of the column to sort ascending or descending), there are 50 cameras with a higher listed QE.

I don't know how to interpret this in light of your statement.

I'm not challenging it, I just don't understand the QE numbers on Bill's site.

 

The Sony A7S is a reference point for full frame cameras. Some of those cameras in that list report QE of greater than 100%, which is impossible!  It demonstrates the difficulty of using DXO results.  Also, if there is any kind of data filtering going on in the those raw files then it will lead to an artificially high value of QE.

 

 

Is this full-system camera QE we are talking about including microlenses, BSI, etc?

Is this QE calculated for the full frame, or just on a pixel basis?

Is the data normalized for pixel size if it's on a per pixel basis?

Yes microlenses will lead to an increase in QE because fewer photons are lost.  QE is simply the ratio of photons converted to electrons compared with the total number of incoming photons.  So it can be calculated per pixel.

 

 

In general conceptual terms, how is QE calculated from DxO data?

I ask, because some there are 6 cameras listed with 100% or higher QE.

I wish I knew how it is done.  It's something I need to look into sometime.

 

 

I've always been skeptical of DxO data because of discrepancies like this. Maybe there are reasons for this that explain it, but I thought I went looking for an explanation years ago and everything was opaque at DxO.  Or maybe I'm just not smart enough to understand it.

Me also, I'm sceptical of relying on DXO data.

 

Mark

[sharkmelley]

A quick update:

• I managed to obtain some real astro-exposures last night between clouds.  They do show the PDAF stripes when pushed hard in post-processing but I haven't yet tried removing the stripes with calibration.
• When talking about sensor self-heating I should have mentioned that when live-view is being used the camera does become quite warm - this is normal behaviour for mirrorless cameras.  But it doesn't become hot when taking a sequence of long exposures.
• I posted a question on DPReview regarding the odd histogram behaviour in the bias frames: https://www.dpreview...s/post/62937140
• Some really good news:  I haven't yet been able to trigger any visible star damage caused by the spatial filtering (I'm still trying!)  Spatial filtering is certainly taking place but it doesn't appear to affect stars or their colours, at least not using my optics.

Mark

Edited by sharkmelley, 30 July 2019 - 02:11 AM.

[sharkmelley]

I have some good news and bad news.

 

First the good news.  I did some experiments with the silent shutter i.e. using the electronic shutter instead of the physical one.  The bit depth remained unchanged at 14bits.  Other cameras typically reduce the bit depth to 12bits with silent shutter.  Silent shutter is great for long focal length astrophotography because there is zero vibration, so it's great news that it retains 14 bits.

 

Now for the bad news.  It does appear that the PDAF banding is a potential issue. I deliberately tried to trigger it by taking short exposures so the back-of-camera histogram peak was well to the left and I deliberately did not use dithering.  Here is a crop of the stack of 30x60sec calibrated exposures at ISO 800 on my f/2.8 Tak Epsilon:

 

[click on image to see it at 100% scale]

 

If look carefully in the shadows you might well be able to see lines of horizontal noise.

 

I applied PixInsight's Superbias process to emphasize the horizontal structures:

 

[click on image to see it at 100% scale]

 

You can see that the spacing of the horizontal bands is precisely 12 pixels, which is the PDAF pixel spacing.

 

When I looked more carefully at the master bias, and master dark I could see that there is actually a difference in the bands and they both differ from the banding in the lights.  So this means calibration will not remove the banding.

 

It now seems that the PDAF banding will need to be the major area of investigation.  I certainly don't believe it will be a showstopper but I need to better understand exactly what is going on and it will probably be necessary to choose shooting strategies to mitigate the effects. For instance:

• Choose ISO and exposure duration to put the back-of-camera histogram peak in the usual 1/4 to 1/3 from the left.  The intention is to sufficiently drown out the intensity of the bands using noise so that it doesn't become visible during stacking and processing.
• Use dithering to spread out the effects of the banding and make it less visible.

Mark

Edited by sharkmelley, 31 July 2019 - 03:29 PM.

[whwang]

Hi Mark,

I suppose the biases were taken in M mode while the darks were taken in Bulb mode. One thing to try is to switch to Bulb mode for biases and see if it behaves more like the darks.

If you only subtract darks without bias subtraction and dark scaling, does the banding reduce?

Cheers,
Wei-Hao

[maxmir]

Post your master bias and dark along with  the sky images. I would Like to have a look

 

Max

[sharkmelley]

Hi Mark,

I suppose the biases were taken in M mode while the darks were taken in Bulb mode. One thing to try is to switch to Bulb mode for biases and see if it behaves more like the darks.

If you only subtract darks without bias subtraction and dark scaling, does the banding reduce?

Cheers,
Wei-Hao

I'll give Bulb mode biases a try but I don't expect it to help.

 

The example above is actually the one where I subtracted just the master dark.  The one with bias subtraction and dark scaling is marginally worse.

 

Post your master bias and dark along with  the sky images. I would Like to have a look

 

Max

I'll do this if I get a chance tonight.

 

Mark

[whwang]

The example above is actually the one where I subtracted just the master dark.  The one with bias subtraction and dark scaling is marginally worse.

 

Then it is probably not because the camera does something different in Bulb and M modes in dark and bias.

 

Did you apply flat fielding?

[sharkmelley]

Then it is probably not because the camera does something different in Bulb and M modes in dark and bias.

 

Did you apply flat fielding?

Agreed.  Since both the darks and lights were taken in bulb mode then Bulb vs M is not the explanation.  I applied flat fielding but the relative amplitude of the bands in the master flat is so small that it is definitely not contributing to the problem.

 

For anyone interested, I've put a raw NEF and 32bit float TIFFs of the averaged bias, darks, flats, lights here:

https://drive.google...TYqcOnZRhHnhofA

 

 

The problem seems to be that the amplitude of the dark stripes in the master dark does not match the amplitude of dark stripes in the master light.

 

A similar problem happens in the masterbias and master dark as you can see here:

 

[click on image to see it at 100% scale]

 

The dark stripe in the master bias is a single pixel wide but it is 3 pixels wide in the master dark.

 

There are some useful contributions from  Jim Kassson, Horshack and pippo27 in this thread:

https://www.dpreview.../thread/4331154

 

They are discussing the Nikon Z7 but it appears to be the same problem.  One interesting assertion is that the issue does not arise in 12bit shooting.  That's certainly an interesting avenue to explore.  For long exposure imaging of deep sky objects it is probably quite safe to reduce bit depth to 12 because the sky glow noise is sufficient to dither the quantisation.

 

Another interesting point raised is that under certain circumstances a sequence of identical exposures may contain some images with banding and some with no banding.  Possibly that is an explanation why my master dark does not calibrate my master light - they perhaps contain different proportions of exposures with banding.  I can certainly investigate this.

 

The rows containing the banding are known in advance and the amplitude of the banding is of the order of 1 digital unit.  I am 100% certain it is possible to remove the banding from master bias frames and master darks.  It is probably also possible to recognise and remove the banding from individual light frames.  If I'm right then I am confident in my ability to write a special Nikon Z6/Z7 process for PixInsight to perform the necessary functionality.

 

But we're jumping ahead here.  I still want to test the possible mitigation strategies of shooting with the histogram peak at the 1/4 position and/or using dithering or using 12bit acquisition.

 

Unfortunately, for the next few days I am really busy with other priorities so I won't make much (if any) further progress for a while.  But if I get a clear night in this new moon cycle (I'm in the UK!) I'll try out those strategies.

 

Mark

Edited by sharkmelley, 02 August 2019 - 02:17 AM.

[t_image]

The rows containing the banding are known in advance and the amplitude of the banding is of the order of 1 digital unit.  I am 100% certain it is possible to remove the banding from master bias frames and master darks.  It is probably also possible to recognise and remove the banding from individual light frames............

I didn't doubt you would think of this,

but such was my initial thought due to the benefit of the PDAF pixels being predictably geometric through the frame and AP works with stacks!

[sharkmelley]

Latest news: 

Dithering works perfectly to remove the visible effects of PDAF banding in a stacked image.  So does 12bit mode (without dithering).  So does exposing with the histogram peak at the 1/4 level on the back-of-camera histogram (without dithering).   I'll show side-by-side examples in due course.

 

So PDAF banding is essentially a non-issue for long exposure stacked astro-photography.

 

Mark

Edited by sharkmelley, 12 August 2019 - 02:40 PM.

[Jerry Lodriguss]

Latest news: 

Dithering works perfectly to remove the visible effects of PDAF banding in a stacked image.  So does 12bit mode (without dithering).  So does exposing with the histogram peak at the 1/4 level on the back-of-camera histogram (without dithering).   I'll show side-by-side examples in due course.

 

So PDAF banding is essentially a non-issue for long exposure stacked astro-photography.

Outstanding!

 

Would you assume that the Z6 is going to be very similar with the PDAF pixels and being able to get rid of them?

 

Thanks for your effort!

 

Jerry

[sharkmelley]

Outstanding!

 

Would you assume that the Z6 is going to be very similar with the PDAF pixels and being able to get rid of them?

 

Thanks for your effort!

 

Jerry

Hi Jerry,

 

Maybe you mean the Z7?  I would expect the behaviour of the PDAF banding on the Z7 to be identical to the Z6.  So the same techniques should work just as effectively.

 

Mark

[Jerry Lodriguss]

Hi Mark,

 

Yes, I meant the Z7.  

 

That's good news, thanks!

 

Jerry

[sharkmelley]

I ought to add that I have good reason for expecting the behaviour of the Z7 to be the same.  It's because the definitive thread on the Z7 PDAF Banding is here: https://www.dpreview.../thread/4331154  and I am finding that my Z6 behaves in the same way, including the amplitude of the PDAF stripes in the image.

 

One thing I should add is that it is quite possible that PDAF Banding might be visible in the shadow areas of a single astro-landscape exposure.  But don't despair because a RawTherapee tool has already been written by DPReview contributor pippo27 and tested here:

https://www.dpreview...s/post/62250409

 

Mark

Edited by sharkmelley, 13 August 2019 - 12:07 AM.

[whwang]

One thing I should add is that it is quite possible that PDAF Banding might be visible in the shadow areas of a single astro-landscape exposure.  But don't despair because a RawTherapee tool has already been written by DPReview contributor pippo27 and tested here:

https://www.dpreview...s/post/62250409

 

 

Can something like this be directly implemented as part of LibRaw, so every astro software that relies on LibRaw can get rid of the banding?

[sharkmelley]

Can something like this be directly implemented as part of LibRaw, so every astro software that relies on LibRaw can get rid of the banding?

I guess it probably could.  RawDigger certainly has some optional Vendor specific features, though I'm not sure if those are implemented in RawDigger itself or in Libraw.  In any case the astro-software would need to call a specific option to switch the extra functionality on.

 

Mark

Edited by sharkmelley, 13 August 2019 - 03:13 PM.

[mmalik]

The bandwagon (... truly 'BAND' wagon); all that remains is deciding how high to jump...

 

 
"One of the biggest mistakes that Nikon made prior to announcing their full-frame mirror-less cameras was to really over hype it. They were using such incredible buzzwords to describe what their new camera was going to be that when it finally arrived the only thing that was incredible was how underwhelming they actually were. Sony, unfortunately, has been doing something similar with executives describing how amazing this new camera is going to be. The problem with describing as such is that we already have pretty amazing cameras on the market now when it comes to just the specifications. In order for Sony to produce something that really stands out now, they're going to have to produce something far beyond their competitors."

Edited by mmalik, 14 August 2019 - 07:12 AM.

[sharkmelley]

The advantage of the Nikon Z6 wide opening for the lens mount can be clearly seen in the following flats taken on my Tak Epsilon scope.

 

Here's the Sony A7S flat:

 

Here's the Nikon Z6 flat:

 

Note how the extreme corners of the Sony flat are pretty much in shadow.  The problem is even worse using it on my Celestron C11.

 

On the Nikon Z6 there is no problem at all in the extreme corners.  Also note that being mirrorless neither camera has any shadowing (or diffuse reflections) caused by the DSLR mirror box and parked mirror.

 

Mark

Edited by sharkmelley, 15 August 2019 - 03:11 AM.