Ad
Ad
Ad
Pages: « 1 ... 9 10 [11]   Bottom of Page
Print
Author Topic: DXOmark ranks DB image quality well below DSLR!  (Read 33618 times)
Panopeeper
Sr. Member
****
Offline Offline

Posts: 1805


« Reply #200 on: May 28, 2009, 09:37:42 AM »
ReplyReply

Quote from: Ray
Fourthly, whilst the DR at base ISO is not significantly worse than that of the D700, I'd be a bit concerned about the waterproofing issue with the 5D2 that became apparent on Michael's recent Antarctic trip.
The DR of the D700 with ISO 200 is 1/2 EV better than that of the 5D2 with ISO 100. If you have to go with ISO 200, the difference is 2/3 EV.

On the other hand, the 5D2 is 1/2 EV better than the 5D (I can compare only ISO 100 until you send me the other 5D shots).

1/2 EV is a lot. On the other hand, the difference between the D700 and the 50D is much greater: about 1 EV; again, the D700 with ISO 200, the 50D with ISO 100. (I don't engage in the speculation about converting pixels in dynamic range.) IMO the combination D700 and 50D is optimal: one for the DR and clean image, the other for the reach.
Logged

Gabor
BJL
Sr. Member
****
Offline Offline

Posts: 5163


« Reply #201 on: May 28, 2009, 10:53:45 AM »
ReplyReply

Quote from: Panopeeper
Is this the very same sensor, which is used in the Sony A900? If so, why does the A900 create only 12bit raw data, while the D3X creates 12 or 14bit?
The answer seems not to be public information.

One proposed explanation is that the A/D convertors do repeated sampling and averaging to reduce the RMS noise level. This goes with the lower frame rate in 14-bit mode (am I right that the D3X has a far lower frame rate in 14-bit mode?)

Another proposed explanation is that the D3X has a modified sensor which bypasses the on-board A/D convertors in 14-bit mode, using off-board 14-bit ADCs instead. Even if that is done, I would expect the on-chip ISO gain amplification to still be used, as that is likely part of the process of transferring the signal from photosite to column bottom. So if anything, amp. noise would then be more dominant, with the A/D noise presumably being less in 14-bit mode.
Logged
Panopeeper
Sr. Member
****
Offline Offline

Posts: 1805


« Reply #202 on: May 28, 2009, 10:01:27 PM »
ReplyReply

Quote from: BJL
One proposed explanation is that the A/D convertors do repeated sampling and averaging to reduce the RMS noise level. This goes with the lower frame rate in 14-bit mode (am I right that the D3X has a far lower frame rate in 14-bit mode?)
I heard this speculation, already regarding the D3. However, there is a problem with it: analysis of the raw data shows, that the 14bit data is the "natural" version and the 12bit data is digitally derivated from the 14bit. The same is true re the D3X.

I think the question is logical: if Sony is making such a sensor, why don't they make use of it?
Logged

Gabor
ejmartin
Sr. Member
****
Offline Offline

Posts: 575


« Reply #203 on: May 29, 2009, 06:27:36 AM »
ReplyReply

Quote from: Panopeeper
I heard this speculation, already regarding the D3. However, there is a problem with it: analysis of the raw data shows, that the 14bit data is the "natural" version and the 12bit data is digitally derivated from the 14bit. The same is true re the D3X.

I think the question is logical: if Sony is making such a sensor, why don't they make use of it?

The D3 is a totally different sensor design, for which it would be natural for 14 to be the native bit depth.  Do you have evidence that this is so for the D300 and D3x?  The RAW files I have for the D300 suggest otherwise; they have no gaps or spikes in the G channel.
« Last Edit: May 29, 2009, 06:30:45 AM by ejmartin » Logged

emil
bjanes
Sr. Member
****
Online Online

Posts: 2825



« Reply #204 on: May 29, 2009, 07:38:47 AM »
ReplyReply

Quote from: ejmartin
What does downsampling do?  Proper downsampling simply removes all spatial frequencies above the Nyquist frequency of the target image, and chops off the portion of the noise spectrum above that frequency.  I downsampled the 50D image to the 40D pixel dimensions, and this is what happened to the noise spectrum:

http://theory.uchicago.edu/~ejm/pix/20d/po...epower-norm.png

Red and Blue as before; Orange is the 50D downsampled with PSCS3 bicubic, black is downsampling with ImageMagick's Lanczos resampling.  The downsampling, especially with Lanczos, quite faithfully reproduces the 40D noise spectrum.  Of course, the downsampling also removed any image detail at spatial frequencies above the Nyquist frequency of the 40D.

However, one needn't have done the downsampling; it was enough to know that the noise power spectra have the same slope to know that the finer resolution camera is not noisier than its lower resolution cousin; one doesn't need to resample to compare noise, one simply needs to level the playing field by fixing a reference spatial frequency and comparing the noise there, much as one can compare MTF's of different lenses at a fixed spatial frequency to see how much detail they render.  This is my interpretation of what DxO does in their "print" tab; they are implicitly choosing a spatial frequency by fixing an output size and doing a mathematically ideal resampling to a reference Nyquist frequency associated to that output size at a standard resolution in dpi (which translates to a given Nyquist in lph).  The resampling is superfluous, all that matters is the comparison at a fixed spatial frequency.

Now, dynamic range is tightly correlated to noise; the technically savvy photographer's working definition of DR is the range of illumination levels having an acceptably large S/N in an image capture.  Since noise is a function of spatial frequency, so is DR (BJL explained this in a somewhat more intuitive way).  Again, downsampling does not increase DR; rather it is changing the spatial scale at which DR is being measured.  DR at a fixed scale largely doesn't care about downsampling, just as the noise of the 50D at a fixed spatial frequency below the target Nyquist didn't change when the image was downsampled.  To compare the DR of two cameras without fixing a common scale or spatial frequency at which to do the comparison, is a largely meaningless exercise.  However, the finer resolution camera need not be downsampled to make the comparison, rather one needs to measure the dependence of DR on spatial frequency.

Thus far in this thread, I have seen no mention of the Sensor+ pixel binning feature of the new Phase One P65 architecture. If one downsamples 4:1 by pixel averaging, the signal to noise improves as the square root of the downsampling ratio, or by 2:1. However, with the Sensor+ technology (which must be done on chip), 4 pixels are combined into one super-pixel which can be read with only one read noise, resulting in an improvement in the S:N of 4:1. Software downsampling would involve averaging of 4 pixels with 4 read noise contributions.

I would think that dynamic range would be improved, but Phase One claims a DR of 12.5 stops for both methodologies, but the ISO is quadrupled in the Sensor+ mode. I would think that the individual pixels would behave similarly whether or not binning is employed and that the ISO required for saturation of the pixels would not change. Does this mean that full well is not reached with Sensor+? How does binning affect the slope of the noise power spectrum?
Logged
ejmartin
Sr. Member
****
Offline Offline

Posts: 575


« Reply #205 on: May 29, 2009, 09:06:18 AM »
ReplyReply

Quote from: bjanes
Thus far in this thread, I have seen no mention of the Sensor+ pixel binning feature of the new Phase One P65 architecture. If one downsamples 4:1 by pixel averaging, the signal to noise improves as the square root of the downsampling ratio, or by 2:1. However, with the Sensor+ technology (which must be done on chip), 4 pixels are combined into one super-pixel which can be read with only one read noise, resulting in an improvement in the S:N of 4:1. Software downsampling would involve averaging of 4 pixels with 4 read noise contributions.

I would think that dynamic range would be improved, but Phase One claims a DR of 12.5 stops for both methodologies, but the ISO is quadrupled in the Sensor+ mode. I would think that the individual pixels would behave similarly whether or not binning is employed and that the ISO required for saturation of the pixels would not change. Does this mean that full well is not reached with Sensor+? How does binning affect the slope of the noise power spectrum?

They may be claiming 12.5 stops; DxO says a bit over 11.5.  I agree that if their binning technology were clean, they should get two stops of DR.  Going by DxO, they seem indeed to be getting about two stops in the DR chart, and that is borne out in the SNR plots where the ISO 400 and Sensor+ ISO 1600 graphs lie almost on top of one another.  Note that the DR improvement is coming in the shadow end, nothing is changing in the photosites, rather it's the readout of the photosites that is changing so full well is unaffected; so I agree that full well will not be achieved with sensor plus.

BTW, binning will help with read noise but has no effect on shot noise; there the expected gain is 2:1 for 2x2 binning, and it doesn't matter whether you do it in hardware or software.
« Last Edit: May 29, 2009, 09:09:10 AM by ejmartin » Logged

emil
BJL
Sr. Member
****
Offline Offline

Posts: 5163


« Reply #206 on: May 29, 2009, 10:05:51 AM »
ReplyReply

Quote from: bjanes
... with the Sensor+ technology (which must be done on chip), 4 pixels are combined into one super-pixel which can be read with only one read noise, resulting in an improvement in the S:N of 4:1.
That sounds right as far as the dark noise arising from transportation, amplification, and A/D conversion, but not photosite noise (dark current?). But these are probably the dominant sources of dark noise, so gaining about two stops of DR at equal ISO speed seems reasonable.

Quote from: bjanes
I would think that dynamic range would be improved, but Phase One claims a DR of 12.5 stops for both methodologies, but the ISO is quadrupled in the Sensor+ mode.
Roughly, quadrupling ISO removes two stops of DR because the signal strength is reduced by 1/4 while the dark noise level stays the same, so it makes sense that the effects on DR of 4-1 binning and quadrupling of ISO balance out. There should be a DR gain at equal ISO speed (400 or higher) with and without binning.

Why does Sensor+ not allow use of the lower ISO speeds? It seems that either the signal path after binning is limited in charge/current capacity to what a full well gives without binning, or the amplifier (charge to voltage conversion, ISO gain etc.) is limited to this DR. The blame cannot come after the amplification and charge to voltge conversion, because the amplification could be adjusted to keep the output within what subsequent components can handle, and it is hard to see 14-bit or 16-bit ADCs in backs this expensive imposing this 12.5 stop DR limit.


So once again, I see evidence suggesting that a major DR limitation is that the DR of the amplifiers is less than that of both the photosites or the ADC.
Logged
Panopeeper
Sr. Member
****
Offline Offline

Posts: 1805


« Reply #207 on: May 29, 2009, 10:44:16 AM »
ReplyReply

Quote from: ejmartin
The D3 is a totally different sensor design, for which it would be natural for 14 to be the native bit depth.  Do you have evidence that this is so for the D300 and D3x?  The RAW files I have for the D300 suggest otherwise; they have no gaps or spikes in the G channel.
All the Nikon CMOS behave mostly the same way; this includes some pecularities as well. The D300 is different only in the handling of the green channels. There are no spikes in the 12bit version of a channel, because the numerical range is divided by exactly four.

The green clipping is not 16383, it is not even a constant value; this is so with the D3 and D3X as well. The red and blue are stretched to 16383, thus the gaps in 14bit mode; when these are mapped over the 4096 range, the spikes show up. The D3 and the D3X do show the division in the green channels as well.

Anyway, the red and blue are more suitable for the observation. See the attached fine histograms from a 14bit and a 12bit file. Note, that the 12bit version shows only one of the green channels (the other looks more weird due to the clipping point's pecularities). The following list of pixel numbers with specific values explain the spikes; I calculated the red one, it perfectly matches with the 10-out-of-11 population of the 14bit red values.

Code:
Red      Blue
2948    2632+    Level 599
3201    1961
3962+    1749
3184    1852
2958    1791
3727+    2331+
2760    1781
2810    1586
2942    1703
3695+    1577
2729    2116+
2734    1761
3608+    1642
2667    1627
2943    1590
3577+    2152+
2658    1775
2606    1685
2545    1660
3534+    1541
2901    2299+
2645    1760
3543+    1746
2547    1778
2557    1501
3620+    2157+
2723    1760
2572    1716
2651    1684
3592+    1440
2671    2116+
2877    1623
3580    1643
1605-    1557
3726+    1332
3792+    1940
2828    1512    Level 635
« Last Edit: May 29, 2009, 10:47:10 AM by Panopeeper » Logged

Gabor
Ray
Sr. Member
****
Offline Offline

Posts: 8908


« Reply #208 on: May 29, 2009, 11:07:52 AM »
ReplyReply

Quote from: Panopeeper
IMO the combination D700 and 50D is optimal: one for the DR and clean image, the other for the reach.

That's what I think also. I just wish Nikon would hurry up and develop a lens similar to the Canon 24-105/F4 IS, but better. I used that lens a lot on the 5D and I appreciate the IS because so much of my photoraphy is without a tripod. The Nikkor 24-70/2.8 is no doubt a sharp lens, but we're into the age of image stabilisation, aren't we? Are you listening, Nikon?  
Logged
Pages: « 1 ... 9 10 [11]   Top of Page
Print
Jump to:  

Ad
Ad
Ad