Nikon finally gets the noise down.

Started Nov 17, 2007 | Discussions
RobertoAvanzi Senior Member • Posts: 1,003
Re: Interesting! A few questions ...

John Sheehy wrote:

RobertoAvanzi wrote:

John Sheehy wrote:

sensor of the same resolution). The D3 collects about 5x as many
photons as the D300 at the same exposure (53500 and 13000,
respectively, at ISO 200).

Nitpicking, maybe, but lets us do correct math. 53500 is VERY close
to 4x 13000. 5x would be 65000.

Maybe. I thought the word "about" covered the inexactitude.

Well, but I said I was nitpicking.

Roberto

 RobertoAvanzi's gear list:RobertoAvanzi's gear list
Canon EOS 70D Canon EF 70-200mm F4L IS USM Canon EF-S 17-55mm f/2.8 IS USM Tokina AT-X Pro 11-16mm f/2.8 DX Sigma 30mm F1.4 EX DC HSM +2 more
OP John Sheehy Forum Pro • Posts: 21,712
Re: Read noise in ADUs greater w/small pixels

ejmartin wrote:

But I'm still confused. Is this the ISO 100 read noise of the FZ50?

The clipped blacks of the FZ50 ISO 100 RAW are about 1.72 ADU. With about a 1.6x factor for non-clipped, that's about 2.7 ADU.

It would seem so since you're quoting DSLR values for that ISO. So
there is an advantage at low ISO. For high ISO presumably the CCD
FZ50 doesn't do well at all compared to the CMOS DSLR's, with read
noise staying about the same in electrons, while the CMOS goes down
dramatically. At high ISO it would seem the FZ50 is going to do much
worse than the DSLR (though I haven't put in the numbers; running out
of re-edit time), since the read noises of the FZ50 and the 1D3 for
example are going to be comparable at ISO 1600, but you have to bin
16 of the former to make one of the latter and thus the FZ50 binned
is going to have rather worse read noise in aggregate.

It's not really "much worse"; it's just a little worse than what CMOS is doing at its best right now. As I explained on the pbase page, the ISO 1600 on the FZ50 is poorly implemented; is has more read noise than ISO 100 or 200 pushed, so using a pushed 1600, the read noise is 44.7 ADU. 44.7/4.06 (the square root of 16.5) = 11.0 ADU, better pixel-level read noise than Sony CCD DSLRs, and only 2/3 - 1.5 stops worse than Canon DSLRs and the D3.

To realize this benefit of small pixels at high ISO, don't you have
to do it in CMOS, and won't that take up quite a bit larger fraction
of sensor space for electronics (16 times in your example) that will
cause a bit hit in QE? It doesn't seem that there is a real-world
exemplar to examine for that scenario.

Yes, it would probably take doing low ISOs with neutral density, and targetting low photon counts to get read noise down at high ISOs on the tiny-pixel cameras.

This is ISO 1600; the benefit for the FZ50 is even better at low ISOs.

Nice demo. I noticed though that the test conditions were f4; that
would put the FZ50 at or about the onset of diffraction, but still at
the optimum sharpness for the image. Unfortunately the FZ50 won't go
beyond f11, but if one is going to scale up the sensor to DSLR
dimensions one will for many applications want to stop down to
f11-f16 (or more for macro). Do you have a sense for what
diffraction is going to do to the clarity of the FZ50 image at
smaller apertures? I suppose it would be easy enough to simulate in
IRIS by convolving with an Airy PSF, or rather deconvolving the f4
PSF and then convolving with Airy for the smaller aperture (which two
steps amounts to a slightly modified PSF in a single step). Or
redoing the test at f11 might begin to show what is going to happen.

But what is it happening to ? The image is what we take pictures for, not individual pixels. The same diffraction will occur on the focal plane, regardless of pixel pitch, with the same lens and f-stop. Coarser pixel pitches simply make it less of a deal, only because the bigger pitches themselves have already thrown away resolution; not because small pitches cause more diffraction. Small pixel pitches have a reputation for demanding more from the optics, and camera stability, but not because it is inherent in the small pixels to waste any of the possible qualities, but only because the potential is so much higher, and therefore so much easier to fall short of. If you have to stop down your lens and those 2 micron pixel pitches go into extreme oversampling, it is not because the 2 micron pixels are more demanding, per se; it is because the analog diffraction has made the maximum contrast between two pixels so low. At that point, your pixel pitch on the large size can only make the image even softer, and on the small end of the pixel pitch spectrum, you get less return on your storage/readout requirements (counting the same number of photons in less and larger pixels requires less bits).

-- hide signature --

John

Led566 Regular Member • Posts: 450
Re: Imagine the technology behind this Nikon sensor design gets...

thw wrote:

They are stuck with Canon sensors that are probably more
expensive overall, while offering little practical value compared to
the now excellent Sony sensors...

Actually, Sony sensors (designed by Sony, NOT Nikon) are still pretty
bad. The new Sony 12 MP APS-C sensor is only very slightly better
than the 10 MP version, vastly inferior to Canon's 10 MP equivalent.

Provided that the D300 has a Sony 12 MPx CMOS, your assumption is actually proven wrong.

Just take a look at the I-R D300 samples and compare them to 40D/5D samples.

OP John Sheehy Forum Pro • Posts: 21,712
Re: Imagine the technology behind this Nikon sensor design gets...

Led566 wrote:

Provided that the D300 has a Sony 12 MPx CMOS, your assumption is
actually proven wrong.

Just take a look at the I-R D300 samples and compare them to 40D/5D
samples.

But then you're comparing two different noise reductions. The D300 has more noise in the RAW capture than Canon DSLRs. It is not new technology in the way the D3 is.

-- hide signature --

John

ejmartin Veteran Member • Posts: 6,274
Re: Read noise in ADUs greater w/small pixels

John Sheehy wrote:

ejmartin wrote:

there is an advantage at low ISO. For high ISO presumably the CCD
FZ50 doesn't do well at all compared to the CMOS DSLR's, with read
noise staying about the same in electrons, while the CMOS goes down
dramatically. At high ISO it would seem the FZ50 is going to do much
worse than the DSLR (though I haven't put in the numbers; running out
of re-edit time), since the read noises of the FZ50 and the 1D3 for
example are going to be comparable at ISO 1600, but you have to bin
16 of the former to make one of the latter and thus the FZ50 binned
is going to have rather worse read noise in aggregate.

It's not really "much worse"; it's just a little worse than what CMOS
is doing at its best right now. As I explained on the pbase page,
the ISO 1600 on the FZ50 is poorly implemented; is has more read
noise than ISO 100 or 200 pushed, so using a pushed 1600, the read
noise is 44.7 ADU. 44.7/4.06 (the square root of 16.5) = 11.0 ADU,
better pixel-level read noise than Sony CCD DSLRs, and only 2/3 - 1.5
stops worse than Canon DSLRs and the D3.

1.5 stops? That's kind of nasty. That level of difference in performance applies not only to the 1D3/D3, but also my 20D... ouch!

To realize this benefit of small pixels at high ISO, don't you have
to do it in CMOS, and won't that take up quite a bit larger fraction
of sensor space for electronics (16 times in your example) that will
cause a bit hit in QE? It doesn't seem that there is a real-world
exemplar to examine for that scenario.

Yes, it would probably take doing low ISOs with neutral density, and
targetting low photon counts to get read noise down at high ISOs on
the tiny-pixel cameras.

What would this hypothetical camera be for, then? To get high ISO performance you then sacrifice low-ISO performance by throwing away photons, or you sacrifice high ISO performance to get resolution at low ISO that you can't use beyond f8 or so.

This is ISO 1600; the benefit for the FZ50 is even better at low ISOs.

Nice demo. I noticed though that the test conditions were f4; that
would put the FZ50 at or about the onset of diffraction, but still at
the optimum sharpness for the image. Unfortunately the FZ50 won't go
beyond f11, but if one is going to scale up the sensor to DSLR
dimensions one will for many applications want to stop down to
f11-f16 (or more for macro). Do you have a sense for what
diffraction is going to do to the clarity of the FZ50 image at
smaller apertures?

But what is it happening to ? The image is what we take pictures
for, not individual pixels. The same diffraction will occur on the
focal plane, regardless of pixel pitch, with the same lens and
f-stop. Coarser pixel pitches simply make it less of a deal, only
because the bigger pitches themselves have already thrown away
resolution; not because small pitches cause more diffraction.

[snip]

My point was that adding pixels beyond resolution for many applications is as wasteful as adding bit depth beyond dynamic range.

-- hide signature --
OP John Sheehy Forum Pro • Posts: 21,712
Re: Read noise in ADUs greater w/small pixels

ejmartin wrote:

John Sheehy wrote:

Yes, it would probably take doing low ISOs with neutral density, and
targetting low photon counts to get read noise down at high ISOs on
the tiny-pixel cameras.

What would this hypothetical camera be for, then? To get high ISO
performance you then sacrifice low-ISO performance by throwing away
photons, or you sacrifice high ISO performance to get resolution at
low ISO that you can't use beyond f8 or so.

I've been told that there are already sensors that pretty much just count photons for all intents and purposes; they can't count 100K, or 50K or even 25K photons, though. They can only count them in small numbers (maye hundreds). So, you could have a camera that only has shot noise, at extreme high ISOs, with existing technology. They are probably made for night surveilance cameras, or military applications.> > > > This is ISO 1600; the benefit for the FZ50 is even better at low ISOs.

But what is it happening to ? The image is what we take pictures
for, not individual pixels. The same diffraction will occur on the
focal plane, regardless of pixel pitch, with the same lens and
f-stop. Coarser pixel pitches simply make it less of a deal, only
because the bigger pitches themselves have already thrown away
resolution; not because small pitches cause more diffraction.

[snip]

My point was that adding pixels beyond resolution for many
applications is as wasteful as adding bit depth beyond dynamic range.

Yes, but there is a bit of a range where subtle benefits can be had with finer pixels. The only theoretical problem with extreme pixel density is storage and transfer and processing of data. In fact, the ideal information captured by a camera would be a list of photons; where they fall on the analog plane of the sensor, and what their wavelengths are. This data would be the most artifact-free source for anything you want to do with the output. It would be an elephant to store and process, though, especially for low ISOs.

The most storage-efficient way to store photon counts is with big pixels and/or digitization ratios that under-represent the number of photons.

-- hide signature --

John

wolfie
wolfie Veteran Member • Posts: 3,866
D3 - really horrible?

I have seen the Nikon D3 sample shot at ISO 25600 (Hockey goal picture) and thought it was pretty good control of chroma noise, comparable to some other DSLRs at 1600 or 3200. - but then it should be good given the size of the photosites on a full frame sensor (and the price).
--
Shoot the Light fantastic

 wolfie's gear list:wolfie's gear list
Canon G7 X II Olympus E-M5 II Nikon 1 AW1 Olympus 40-150mm F2.8 Pro Olympus 12-40mm F2.8 Pro +1 more
PIXmantra Senior Member • Posts: 1,637
Yes, see by yourself...

...and there is nothing wrong with it, as such speed is just way, way beyond the capabilities of the sensor:

http://www.pbase.com/feharmat/image/89084939/original

At ISO6400, it is clear that it has been maxed out (although still manageable, though):

http://www.pbase.com/feharmat/image/89206710/original

It is what it is.

PIX

-- hide signature --

TIP: If you do not like this post, simply press the 'COMPLAINT' button. Mommy/Daddy are just one click away.

OP John Sheehy Forum Pro • Posts: 21,712
Re: D3 - really horrible?

wolfie wrote:

I have seen the Nikon D3 sample shot at ISO 25600 (Hockey goal
picture) and thought it was pretty good control of chroma noise,
comparable to some other DSLRs at 1600 or 3200. - but then it should
be good given the size of the photosites on a full frame sensor (and
the price).

The balance between luminance and chroma noise on any two RGB cameras with CFA filters depends on the conversion routines. In the unlikely event that one camera leans more towards luminance noise, it is because there is a large low-frequency component to the noise, and that would not be good.

-- hide signature --

John

Ominous Veteran Member • Posts: 4,304
Detail GONE at 3200 iso

Wow. I need to try that on my 5D and see if small detail vanishes like that.
--
http://www.pbase.com/ewhalen

PIXmantra Senior Member • Posts: 1,637
Quite puzzling, isn't it?

Ominous wrote:

Wow. I need to try that on my 5D and see if small detail vanishes
like that.

...Obviously, we can not conclude with 100% certainty that such detail is not in the RAW files but... it is not there, for sure, on the .JPGs.

A tale-story of Nikon's on-board noise-reduction, though.

I would also need to test the 1DMKIII in a similar situation, although the preliminar ISO3200 shots I have seen do not exhibit such loss of low-frequency detail.

Time will tell.

PIX

-- hide signature --
-- hide signature --

TIP: If you do not like this post, simply press the 'COMPLAINT' button. Mommy/Daddy are just one click away.

Amin Sabet
Amin Sabet Veteran Member • Posts: 6,763
Re: Detail GONE at 3200 iso

Looks clearly better than my 5D at 3200+ ISOs based on these and other samples. Time will tell, but that is my impression. IMO, Lightroom makes a mess of high ISO files in general.

 Amin Sabet's gear list:Amin Sabet's gear list
Sony a9 Panasonic Lumix DMC-GX85 Nikon D500 Panasonic Lumix G 14mm F2.5 ASPH Panasonic Lumix G Macro 30mm F2.8 +11 more
PIXmantra Senior Member • Posts: 1,637
NOPE...

Amin Sabet wrote:

Lightroom makes a mess of high ISO files in general.

...It just shows them how they are, based on the spatial response/bias inherent to Lightroom's RAW-to-RGB conversion math.

Nothing else, nothing more.

-- hide signature --

TIP: If you do not like this post, simply press the 'COMPLAINT' button. Mommy/Daddy are just one click away.

Amin Sabet
Amin Sabet Veteran Member • Posts: 6,763
Re: NOPE...

...It just shows them how they are, based on the spatial
response/bias inherent to Lightroom's RAW-to-RGB conversion math.

Right. While doing exactly that, it makes a mess out of high ISO files. I greatly prefer the spatial response/bias inherent to C1's RAW-to-RGB conversion math.

 Amin Sabet's gear list:Amin Sabet's gear list
Sony a9 Panasonic Lumix DMC-GX85 Nikon D500 Panasonic Lumix G 14mm F2.5 ASPH Panasonic Lumix G Macro 30mm F2.8 +11 more
PIXmantra Senior Member • Posts: 1,637
That's because...

Amin Sabet wrote:

...It just shows them how they are, based on the spatial
response/bias inherent to Lightroom's RAW-to-RGB conversion math.

Right. While doing exactly that, it makes a mess out of high ISO
files. I greatly prefer the spatial response/bias inherent to C1's
RAW-to-RGB conversion math.

...C1 filters the hell out of the low-frequency spectral component of noise, and that helps a lot.

In any case, I do not disagree with the improvement in the "look", but I must warn you that these is what these images are made of.

Simply turn both Luma and Chroma noise reduction to ZERO on C1 and you will see.

-- hide signature --

TIP: If you do not like this post, simply press the 'COMPLAINT' button. Mommy/Daddy are just one click away.

Amin Sabet
Amin Sabet Veteran Member • Posts: 6,763
No, it's not.

PIXmantra wrote:

Simply turn both Luma and Chroma noise reduction to ZERO on C1 and
you will see.

I do that every day. You really think someone would use C1 to process files day after day and never try setting the NR to the minimum setting? It still looks far better than Lightroom. Lightroom structures high ISO noise into little diagonals, the minimum luma NR setting in LR still looks like excessive luma NR to me, and the chroma NR in LR is not effective with coarse color noise. My opinion differs from yours, but that's not due to missing some setting in C1 or LR.

 Amin Sabet's gear list:Amin Sabet's gear list
Sony a9 Panasonic Lumix DMC-GX85 Nikon D500 Panasonic Lumix G 14mm F2.5 ASPH Panasonic Lumix G Macro 30mm F2.8 +11 more
PIXmantra Senior Member • Posts: 1,637
I will try to post some samples...

Amin Sabet wrote:

PIXmantra wrote:

Simply turn both Luma and Chroma noise reduction to ZERO on C1 and
you will see.

...At ISO1600, between C1 3.7.7 and Lightroom, from the exact same file, so I can explain myself better.

PIX

-- hide signature --

TIP: If you do not like this post, simply press the 'COMPLAINT' button. Mommy/Daddy are just one click away.

Amin Sabet
Amin Sabet Veteran Member • Posts: 6,763
Re: I will try to post some samples...

PIXmantra wrote:

...At ISO1600, between C1 3.7.7 and Lightroom, from the exact same
file, so I can explain myself better.

Looking forward to it. In the meanwhile, I'd like to hear your thoughts about some ISO 800 crops from a new compact camera, the Ricoh GRD II. Do you think these crops -> http://forums.dpreview.com/forums/read.asp?forum=1013&message=25783894 suggest that Ricoh is doing more NR on the RAW files of the GRD II as compared with the GRD? The RAW files themselves have been kindly provided by Pavel Kudrys here -> http://ricohforum.com/phpbb/viewtopic.php?f=39&t=472

 Amin Sabet's gear list:Amin Sabet's gear list
Sony a9 Panasonic Lumix DMC-GX85 Nikon D500 Panasonic Lumix G 14mm F2.5 ASPH Panasonic Lumix G Macro 30mm F2.8 +11 more
PIXmantra Senior Member • Posts: 1,637
Here you go (ISO3200 samples LR-vs-C1 3.7.7)...

...Left is Lightroom with custom calibration and maximum possible dynamic range (all withing Lightroom controls), and right is C1 3.7.7. All from 1D3 samples, and are particularly rich in LOW frequency detail and harshly lit, which is was intented in all purpose:

http://www.pbase.com/feharmat/image/89578837/original
http://www.pbase.com/feharmat/image/89578844/original
http://www.pbase.com/feharmat/image/89578847/original
http://www.pbase.com/feharmat/image/89578851/original

NR settings for LR where (L=20-25, Chroma=12-13), and C1 was (L=0, Chroma=10-12). Additional noise reduction in C1 is L-channel destructive.

Sharpening for LR was (20%,0.7,25,15-20), and for C1 was (0,50).

No further noise reduction nor anything else performed outside of each converter, in any form or shape, other than the vis-a-vis capture, and conversion from EIZO CG241W profiled monitor to sRGB (Eizo's space is almost AdobeRGB, upon closer examination of its profile in GamutVision).

Enjoy!

-- hide signature --

TIP: If you do not like this post, simply press the 'COMPLAINT' button. Mommy/Daddy are just one click away.

Amin Sabet
Amin Sabet Veteran Member • Posts: 6,763
Re: Here you go (ISO3200 samples LR-vs-C1 3.7.7)...

Thanks for the samples. I don't know the technical terms to properly describe the characteristics of LR high ISO conversions that I find distasteful, but your examples are similar to what I see when I use LR for high ISO. The way noise and fine details are rendered by LR at high ISO is just not pleasing to me. As for the loss of information in the circled regions of screenshot 3 for C1, I'd like to have a try at processing that file for myself. I use C1 to process high ISO 5D files all the time, and that doesn't look right to me.

As an aside, I think the minimum luma NR setting on C1 3.7.7 destroys more detail than the zero luma NR setting in C1 4 beta 2, though either looks better to me than LR for high ISO files.

 Amin Sabet's gear list:Amin Sabet's gear list
Sony a9 Panasonic Lumix DMC-GX85 Nikon D500 Panasonic Lumix G 14mm F2.5 ASPH Panasonic Lumix G Macro 30mm F2.8 +11 more
Keyboard shortcuts:
FForum MMy threads