DR, DxO, DSLR, MFDB, CMOS, CCD

[ErikKaffehr]

Ray,

If, if, if

Lloyd Chambers havs an excellent review of the S2 on his DAP site. It's a pay site, in my view worth every penny. Lloyd published a couple of DNG images from the Leica S2 and the Nikon D3X. unfortunately on quite restrictive terms. I have played a bit with those images and came up with a few interesting observations. The key observation is that although the S2 images have without any doubt higher resolution the Nikon images are smoother and have much less artifacts. So I essentially prefer the Nikon images but the look of the images is much dependent on the processing parameters.

Best regards
Erik

Hi Bart,
I don't think I've misconstrued your intentions here. Best of luck!   . This issue is particularly paradoxical when one takes into consideration that comparison that Michael made, at A3+ print size, using a P45+ and the 15mp Powershot G10.  If one looks at the DXO scores comparing those sensors at the normalised print size of  8"x12", the P45+ is way ahead with 1/4 the noise at 18% grey, 4.7 bits better color sensitivity and 2 bits better tonal range, yet at double the size of the DXO normalized print, experienced photographers were not able to distinguish any significant qualitative difference between the two prints, except the shallower DoF in one of the prints which was an indication that that was the print made from the P45 image.

If Michael had equalized the DoF in both shots, using F18 or F22 with the P45+ instead of F11, we might have had the even more amazing situation whereby the G10 print was confused as orininating from the P45 camera because it appeared slightly sharper.

It's because such differences can go unnoticed in a 'real world' print (as opposed to a print that is specifically designed to highlight such differences, such as a resolution line chart which highlights resolution differences), that I'm very skeptical that relatively small differences in 'effective' lens MTF performance, due to sensor size & pixel pitch, could have any significant bearing on the perception of DR.

When I first visited the DXO Mark site, I was a bit confused about the apparent conflict between SNR and dynamic range. We tend to associate a good SNR figure with a high DR, yet there are many examples of sensors having a worse DR but a better SNR on the DXO website. I presume this situation arises because DXO only measure SNR at 18% grey, which is a fairly light shade of grey, in the mid-tone range. I presume, if DXO were to measure SNR at significantly darker shades of grey, the SNR figures would correspond more closely with the DR figures, ie. a higher DR would tend to be associated with a higher SNR at the darker shade of grey. Is this correct?

Since DXO Mark don't provide results at larger, normalised print sizes, such as 23"x31" for the P65+ at 300 dpi, I can only speculate how the DR of an interpolated D3X image would compare, at this size.

Since the D3X image has a 1.33 stop DR advantage at the pixel level, and a 2/3rds stop advantage at the downsampled size of 8"x12", I think it would be reasonable to guess that at any larger, interpolated print size, the DR of the D3X print would not be worse than that of the P65+, or at least not perceptibly worse, because the resolution advantage of the P65+ image, in those darker shades of grey in the 11th and 12th stops, probably doesn't exist. It's been obliterated by noise.

If the resolution advantage in the darker shadows hasn't been recorded by the P65 sensor in the first instance, then it cannot magically materialize, however big the print size. Nor, I would imagine, could any slight advantage from an effectively higher MTF from the lens provide that additional detail in areas of the image that are swamped in noise.

However, I would of course expect to see higher resolution in the 23"x31" print from the P65, in the range from the lower mid-tones to the highlights. I would expect to see sharper eylashes and a slightly smoother, creamier complexion of the model's face. But I wouldn't expect to see a more detailed chihuahua, hiding in the dark, under the couch.

[Ray]

The downsampling algorithm has an effect on SNR improvement. For e.g., is it binning (box filter), or sinc, or lanczos, or something else. In general the implicit assumption has been that pixels are simply binned/averaged, derived from all those discussions on smaller/larger sensor SNR behavior. However, in practise, one would use binning seldomly for downsampling as more sophisticated methods are used. The good news is that it is possible to derive an analytical, closed form, solution to SNR improvement in a generalized downsampling operation that can cater to different methods, say lanczos downsampling.

But the point is not to go into detail of such a solution, but just to emphasize that quoting a figure such as '2/3rds stop advantage at the downsampled size of 8"x12"' without mentioning how the image was downsampled is not full piece of information.

Joofa

Joofa,
Yes, I know that different methods of downsampling and interpolation can affect the result. But what can I do about it? I'm not aware of anyone having the full information on any topic or issue under the sun. There are always some unanswered questions, approximations, assumptions or uncertainties.

As regards a dynamic range comparison between the D3X and P65+, the DXOmark results are the only 'substantiated' results I'm aware that exist. I'm therefore not able to quote any other results whether they may be more accurate, complete or incomplete.

I understand that DXOmark results are based on an analysis of the RAW file from each camera they test, thus avoiding endless dispute about which RAW converter may be best for which camera.  I don't know how DXO derive the 8x12" normalised image. Perhaps those DR figures relating to the downsampled images are merely mathematical calculations based on the original analysis of the sensor at the pixel level.

I'm also aware from comparisons of my own images that downsampling an image tends to reduce noise, so the DXOmark results for their 8mp 8"x12" normalised images tend to be credible for me.

[Ray]

Ray,

If, if, if

Lloyd Chambers havs an excellent review of the S2 on his DAP site. It's a pay site, in my view worth every penny. Lloyd published a couple of DNG images from the Leica S2 and the Nikon D3X. unfortunately on quite restrictive terms. I have played a bit with those images and came up with a few interesting observations. The key observation is that although the S2 images have without any doubt higher resolution the Nikon images are smoother and have much less artifacts. So I essentially prefer the Nikon images but the look of the images is much dependent on the processing parameters.

Best regards
Erik

Erik,
I've noticed the other interesting thread on the Lloyd Chambers' review of the Leica S2, but I'm not subscribed to that pay site. I'm not sure I need to see in-depth reviews of equipment I'm never likely to buy. At the moment I feel as though I have enough photographic equipment but not enough time to use it thoroughly and experiment with its potential. Nevertheless your report on your comparison of the DNG images you downloaded is interesting.

As regards equipment purchases, the next logical step for me would be a 5D2. My highest resolution camera is still the 15mp 50D. If I hadn't previously purchased a Nikkor 14-24/2.8, which led to the purchase of a D700, I would probably by now have a 5D2 and one of the new 17 or 24 TSE lenses (or perhaps both).

On my recent river cruise in Europe and Russia, I frequently carried around my neck both the D700 with 14-24, and D50 with 17-55. I must have got asked about 50 times, often by complete strangers, why I was carrying two cameras.

I did my best to explain that I effectively get a high quality F2.8 zoom from 14-88mm and that no such single F2.8 zoom lens of similar quality and range exists anywhere at any price.

When I've got time, I'd like to do a thorough comparison of these two camera and lens combinations, comparing the Nikkor at 24mm with the Canon at 17mm (which is effectively 27mm on FF).  My impression so far is that the advantages of the greater DoF of the cropped format, combined with the advantages of the image stabilisation of the Canon EF-S 17-55/2.8, sometimes result in better image quality, or at least not worse image quality.

For example, shooting a fresco in a church where flash and tripod is not allowed, I might use the D700 at ISO 1600 and F2.8 but can use the 50D at ISO 800 or even 400 at F2.8 because of IS, which the Nikkor lens doesn't have. When photographing something more 3-dimensional in low light, when the DoF at F2.8 is too shallow on full frame, I might need to use F5 at ISO 1600 with the D700, as opposed to F2.8 at ISO 200 with the 50D, on the basis that IS provides a 2-stop shutter speed advantage and the smaller format provides a 1.6 stop DoF advantage. In such circumstances, there's no doubt that the 50D at ISO 200 has lower noise and higher DR than the D700 at ISO 1600. On the other hand, there's no doubt that the Nikkor 14-24 at 24mm and F5 is sharper than the Canon 17-55 at 17mm and F2.8.

Cheers!

[joofa]

Joofa,
Yes, I know that different methods of downsampling and interpolation can affect the result. But what can I do about it? I'm not aware of anyone having the full information on any topic or issue under the sun. There are always some unanswered questions, approximations, assumptions or uncertainties.

I'm also aware from comparisons of my own images that downsampling an image tends to reduce noise, so the DXOmark results for their 8mp 8"x12" normalised images tend to be credible for me.

Hi Ray,

I know what you are saying and I agree with it. SNR would typically improve in downsampling operations. The reason is that in this case the signal has a correlation structure where as the noise by assumption doesn't.

I just wanted to point out that according to the link that Bart provided a few messages up DXO also assumes a "binning-like, fit-to-size" operation in their "Print" DR comparison. Where as, in reality, hardly any serious resampling operation would use binning. (That is another topic that in the first place I personally think that even the usual binning SNR improvement is not fully correct as done by DXO, and others, but that is a much debated topic by now).

Regards,

Joofa

[Guillermo Luijk]

Now, many experienced observers clearly see a 4-6 stop advantage with MFDBs over DSLRs. I don't have any issue with that, but I cannot understand where it is coming from.

It is coming from their lack of knowledge about what DR is.

Typical useful per-pixel DR in DSLR's is 8-9 stops (SNR criteria of 12dB, which can be considered a good reference for photographers). A back with 6 extra stops over that, i.e. 14-15 stops of DR, could virtually capture 99% of HDR scenes in real life with a single shot and without noise problems in the deep shadows. Obviously this is not the case.

This kind of scene is one of the most challenging for any digital camera:



It has about 15 stops of DR:




Has anyone with a back tried to shoot in a dark room with a sunshine outdoor window, and was able both to preserve the highlights in the window and display no visible or very acceptable noise in the deep shadows inside? if a back does this then I'll believe it has 6 stops of advantage over my 350D's DR.

A 15 stops DR camera is very far from what can be found in the market today so the 4-6 stops advantage claim is nonsense.

Regards

[BJL]

It is coming from their lack of knowledge about what DR is.

Typical useful per-pixel DR in DSLR's is 8-9 stops  (SNR criteria of 12dB...) ... A back with 6 extra stops over that, i.e. 14-15 stops of DR ...

Thanks. And it is even more obvious, since the MF sensor and back makers have given as a hard 12-stop upper limit (engineering SNR, criteria of 0dB) via their spec. sheets. The only way for high end DSLR's to be six stops worse would to start require them being down at about 6 stops "usable" DR. How high a SNR threshold would be needed to justify claiming that low a DR for DSLR's?

[Bart_van_der_Wolf]

Typical useful per-pixel DR in DSLR's is 8-9 stops (SNR criteria of 12dB, which can be considered a good reference for photographers).

Hi Guillermo,

The 'typical' 12 dB criterion is a bit arbitrary IMHO, for some it's too noisy already, and for others there is a lot that can still be done with decent noise reduction. Besides, there are already differences in Raw converter behavior with all noise reduction zeroed out. Obviously, it is also about a single Raw file, not e.g. a 'Zero-noise'ed or HDR file.

Just trying to understand what you are saying, but as I understand it it is more of an example than a statement, correct?

Cheers,
Bart

[Guillermo Luijk]

Just trying to understand what you are saying, but as I understand it it is more of an example than a statement, correct?

Just wanted to point that a 4-6 stops advantage in DR is too huge to be true, no matter what SNR criteria we use (roughly 6 extra stops means 64 times or 36dB higher SNR in the deep shadows). There is not such difference even between a compact camera and a FF DSLR, where relative differences in sensor size are much larger.

Yes 12dB is an arbitrary criteria but surely will satisfy many more photographic applications and users than 0dB.

The only way for high end DSLR's to be six stops worse would to start require them being down at about 6 stops "usable" DR. How high a SNR threshold would be needed to justify claiming that low a DR for DSLR's?

Without trying to be very precise, I did some SNR measurements for the Canon 5D, 7D and 5D2. Looking at this per-pixel SNR plots:



To justify a DR of 6 stops, on a Canon 5D2 at ISO100 we must rise up to a SNR=27dB!
BTW according to that plot, the per-pixel DR(0dB) for the Canon 5D2 would be 11 stops (DxO says 11,16 stops).

Regards

[EricWHiss]

...

[ErikKaffehr]

Hi,

I just looked at a couple of well exposed files from Leica S2 and Nikon D3X courtesy of Lloyd Chambers, when I increased exposure in Lightroom to +3.6 and added fill light 35% the Leica S2 image was clearly worse than the D3X, so I couldn't see any increase in DR in the dark end. Regarding highlights I don't know.

Best regards
Erik

Thanks. And it is even more obvious, since the MF sensor and back makers have given as a hard 12-stop upper limit (engineering SNR, criteria of 0dB) via their spec. sheets. The only way for high end DSLR's to be six stops worse would to start require them being down at about 6 stops "usable" DR. How high a SNR threshold would be needed to justify claiming that low a DR for DSLR's?