Definitive A900/D3X noise/ISO samples

[mdcromer]

Don,

All full sized images are on my site under the ++ symbol. BTW the links you posted will turn into a pumpkin at midnight. . . Use the "permalink" icon for permanent links to my images, the gallery links shuffle around once per day. . . Also about 1/2 were from other cameras than the R1 such as Sony 828, 717 which are a completely different ball of wax. . . I shoot with a tripod most of the time and base ISO, very rarely I shoot over base ISO and/or handheld. . .

The images were not processed for low noise -- for me, detail limits all 10-12MP detailed landscape images to 12 inches on the long side for best results, 18 absolute max regardless of noise.

24MP is enough for excellent 18" prints, good to 24".

[alphaomega]

Here is an interesting contribution from bob2 on the Canon 1D/5D thread

http://forums.dpreview.com/forums/read. ... e=30449425

The concluding part concerning mostly a comparison of the A900 with D3X sensor (and other common Sony sensors) performance is quite interesting and perhaps disturbing if at all true. I personally find it hard to believe though:

* The performance of these sensors in Nikons is so much better and Sony's in and out of camera development software so dire that one could almost put forward a convincing case that Nikon must have Sony Semiconductor in some kind of unbreakable contract (presumably made prior to Sony's acquisition of KM's SLR interests) that forces Sony Imaging to artificially hobble the performance of their cameras.

Maybe he has not got to the idea of DK & Sonolta's point about Nikon's black clipping and Sony's strong emphasis on lower ISO performance at the expense of higher ISO results. I would agree with him that IDC could do with some major improvements including the inclusion of CA removal.

It is also worthwhile to read this additional thread from bob2:

http://forums.dpreview.com/forums/read. ... e=30451444

Interesting views about how the Sony wafer could be completed by Nikon in their own facility.

[fastson]

That doesnt make sense. Arent the A700 and the D300 using the same sensor, and very comparable now with V4?

[David Kilpatrick]

I have been doing some unscientific but meaningful tests tonight.

Here's the score: give two exposures of 0.4 seconds at f/11, and the Nikon at ISO 100 matches the highlight of the Sony at ISO 160 (the clipping level highlight - visible detail without using minus correction in raw processing). This is using the same transfer curve for each camera. But the Sony has mid-tones and shadows which are, as you would expect from the higher ISO, 2/3rds of a stop brighter.

It is very clear how Nikon achieves the lower noise, it's exactly the same reason they achieve the lower ISO - but don't get the same highlight rolloff. They simply call what would be approx ISO 160/200, ISO 100; they use less gain in the very early stages. Easy. You give more exposure, less need to amplify the photon count. Highlights get a slight loss (to get the same highlight information, you must underexpose the Nikon by 2/3rds of a stop, but then of course you get a rather dark picture in the rest of the tones).

It is SO simple as a concept because the raw processor software writers must then use a flatter curve for Sony's raw, a lighter midtone curve for Nikon, and for whatever reasons this is introducing less noise (final stage gain boost, in effect) than the in-camera equivalent. This is based on using the Nikon curve written for RAW Developer, to process Sony files. If you do so, Sony's 3200 becomes 5000 (or Nikon's 6400 becomes 4000,whichever way you see it). Everything I can see points to the sensors being practically identical, and the alternative approaches to in-camera gain curve versus post-processing gain curve (combined with different 'ISO speeds') are what make the difference.

I have also briefly tested ISO 3200 Nikon 12-bit lossless compressed versus 14-bit uncompressed - no substantial difference in the noise type or level, and as far as I can tell, absolutely no point in using 14-bit!

David

[mdcromer]

Not sure that I follow you 100% David. But sounds like you might be on to something. . .

Looking forward to seeing the "part 2" article on the noise comparison where you elaborate on these latest comments with some visual aids and some pictures.

It's good to know that the 14 bit vs. 12 bit mode isn't making any difference.

[David Kilpatrick]

It is really difficult to explain. I need to capture a video clip from the screen showing the adjustments needed to demonstate what this is about. But I understand now exactly what Nikon is doing - if you apply ANY amplification to sensel-location based data before it is written to raw, you get hot pixels, coloured noise etc. If you apply more amplification to de-Bayered data, you are averaging the noise amplified across the deBayer convolution (the pattern of sensels which is used to create a single image RGB image pixel). This is Nikon's strategy and it also explains why they expand the image data to fit the 0-255 deBayered levels straight out of the box.

What is surprising is that this strategy means doing less in the camera. It's very simple despite being being hard to explain in non-technical terms (or, in technical terms, for me - I do not know the correct words to describe what seems to be happening).

David

[David Kilpatrick]

Let me explain. The Nikon is underexposing the sensor, which means less highlight detail gets clipped. It can then apply a tone curve which lifts the bulk of the data slightly higher up the scale then usual, and apply a more gentle roll off in the highlights. This would explain neatly the D300's superior highlight dynamic range measured in our review compared to the A100 (4.1 EV vs 3.3 EV).
--
Andy Westlake

This is fairly old now and probably does relate to the D300/A100, but the strategy with the D3X is to do this and then restore the normal exposure! Really, the D3X is underexposing an ISO 200 native setting (ISO 160) or would be if the camera was set to 200/160. Instead, they set it to 100. I can not compare the D300 and A100 now, the present situation is not identical, but the methods are the same - Nikon is manipulating sensor gain, ISO ratings, and post-processing curves to maximum effect. The trade-off in the D3X is a loss of highlight roll-off, no matter what the D300 had. Sony is now using an in-camera curve which lifts the data values. I don't actually think Andy got it right because unlike RAW Developer, the converters he was using did not give access to the tone curve used when converting the raw. So he was not seeing some hidden processing during raw conversion (out of the camera).

David

[stevecim]

With my totally untrained eye, the biggest differance I see is all the tiny red pixels in the black areas.
Yellows, green,blue, red and light gray look OK, but blacks, dark grays, dark greens get blown away with all the red dots, the a350 gets the red dots in black areas and black dots in red areas

But that's just my newbie view

Stephen

[j_p_p]

check this RAW comparison :

http://www.imaging-resource.com/PRODS/E5D2/E5D2RAW.HTM" onclick="window.open(this.href);return false;

[mdcromer]

Don,

What David is claiming is that you should be able to get similar results from the Sony by turning down the ISO and then underexposing 1, 2, 3 or more stops, and then bring them up in post processing.

He is claiming that this is the strategy that Nikon is using (less on-chip amplification and more tone curve RAW conversion amplification).

Makes perfect sense if the on-chip amplification is the source of the noise issues.

Maybe reshoot one of the prior tests at ISO 800 on the Alpha 900 and underexpose 3 stops, then lift in post? That will mean a lot less on-chip amplification and a lot more RAW convertor amplification.

I think Emil was recommending this with the Alpha 900, although he went as far as to recommend shooting at the base ISO, and getting all your ISO boosting in post.

[douglasf13]

David K, this is a very, very pertinent finding that should be explored. It echoes Iliah Borgs findings. I can't wait to see some samples of the a900 win the D3x curve. Good work.

Sonolta, your machine gun posting hasn't shown that you understand the concept David is trying to explain. He is essentially saying that Nikon uses less amplification in the camera than the A900 and boosts with a tone curve in the RAW processing. Like I said in an earlier post, this is a trade off with the nice a900 highlight roll off.

[David Kilpatrick]

Don,
What David is claiming is that you should be able to get similar results from the Sony by turning down the ISO and then underexposing 1, 2, 3 or more stops, and then bring them up in post processing.
He is claiming that this is the strategy that Nikon is using (less on-chip amplification and more tone curve RAW conversion amplification).

To underexpose the Sony, with the internal camera workings as they are now, might produce more noise. There is some argument for just using the Sony at ISO 800 max, and underexposing if you want 1600 or 3200, then pulling the result back in raw conversion. I know that underexposing ISO 400 by -1 or -2 works well but that only gets you to 1600 max.

Nikon's approach is turn the 'read' values down by 2/3rds of a stop (presumably by clipping the shadows, discarding low values), flatten the midtone/shadow boost given later in the pre-raw processing, and allow 1 stop more exposure by assigning a lower ISO rating for metering purposes. Nikon ISO 100 = Sony ISO 200 when both are compared for highlight exposure not shadows. This results in very clean noise-free data to which a raw conversion curve can be applied safely. They may also be doing some NR to the pre-raw data.

If you overexpose the Sony by one stop, then pull in the result in raw processing, you might get some of the benefits but you can never imitate the total removal of low value shadow data which is responsible for so much noise.

It is easy to see why this works well at low ISOs and rather harder to see how it is benefiting high ISOs. Tests already show that Sony is overstating high ISO values in the A900, and practical results show that treating 6400 as if was really 3200 can transform the result. That would be fine if the Nikon implementation did not go up to 6400 as well, but it does. Nikon is doing more useful processing somewhere along the line, or their speed range would be 50 to 3200 with no boost to 6400.

David

[01af]

What David is claiming is that you should be able to get similar results from the Sony by turning down the ISO and then underexposing [...] He is claiming that this is the strategy that Nikon is using ...

No, just to the contrary! David said, Nikon's strategy is to deliberately over-expose by approx. 2/3 of an f-stop. This naturally will reduce noise in the dark tones and midtones but will also cause more clipping in the highlights.

Furthermore, raw data always needs a gamma curve to be applied, or the midtones will look way too dull and dark (no matter whether you're under- or over-exposing a bit). Nikon's strategy is to cut away the noisy darkest tones (which they can afford due to the over-exposure) and then to leave lifting the midtones to the software after demosaicing (which will distribute the remaining noise, similar to a dithering algorithm).

Sony's strategy, if I understand correctly, is to give less exposure, to do no shadow clipping, and then to apply some of the midtone lifting on-chip, before demosaicing. The benefit is finer tones across the tonal range, a wider dynamic range, and a film-like highlight roll-off, at the expense of more noise.

We can also put it this way: the analog value of what ends up in the final RGB image as a midtone gray, i. e. RGB (128/128/128), is lower in the Sony sensels and higer in the Nikon sensels, before A-to-D conversion. Both approaches do make sense; both have their strengths and weaknesses. The problem (for Sony) is this: to fully appreciate the strengths of Sony's approach, it takes looking at prints, not pixels. However, pixel-peepers prevail these days, so most people don't understand how sensible Sony's approach really is.

-- Olaf

[Greg Beetham]

I'm afraid I still don't get it, if Nikon is over-exposing by 2/3ish- stop....how come they are using faster shutter speeds?? with the same settings everywhere else??
Greg

[Greg Beetham]

I'm just thinking out loud here..... what's the chances of Nikon being able too vary the gain selectively across the sensor?? such as potentially overexposed areas receiving less amp gain??
Greg
ps. It's Christmas morning here, now, so season's greetings everyone.