ejmartin
Offline
Upload & Sell: Off
|
 p.2 #5 · Dynamic Range Measure | |
jkurkjia wrote:
ejmartin wrote:
The figure in that article which has to do with actual DR is figure 5. DR plotted there is the so-called "engineering" DR, which is raw saturation level (max possible raw value of a pixel patch) over read noise (min possible raw value of a pixel patch), both measured at a fixed ISO. This is a rather liberal definition of DR, but has the advantage of being defined in terms of raw sensor data and properties, unfiltered through any raw converter.
The Figure 5 DR curves for the 1DmkII and 1DmkIII at low ISO are more in line with what Canon stated in the 1DmkIII white paper. Without quantifying the DR, Canon stated that the DR of the two cameras (1DmkII and 1DmkIII) is about the same at low ISO speeds.
It's very peculiar "to me" that Canon Marketing chose not to say anything positive about "high ISO DR"; they just kind of stopped after the low ISO commentary. You know, it's just not like marketing to stop short of saying anything positive if in fact there is something positive to say in the first place. In fact, even if there isn't anything positive to say marketing will usually figure out some way to create a positive spin without actually lying.
Or sometimes they just lie, like that laughable sunset shot showing the superiority of 14-bit files. But as in politics sometimes the Big Lie works well.
I always figured (maybe "guess" is a better word) in-camera noise reduction was performed at the high ISO settings for the 1DmkIII prior to creation of the RAW data; maybe, just maybe "if" the in-camera noise reduction also wipes out some data, then it would be illegitimate to actually claim an increase in DR at the higher ISO speeds. Do you have any thoughts on this subject? Do you believe the Clarkvision DR numbers at high ISO speeds is legitimate for the 1DmkII and 1DmkIII bodies?
Regards,
Joe Kurkjian
I haven't seen any evidence of NR on raws at any ISO in the 1D3.
There is something a bit odd in the data quoted for the 1D3 on Clark's summary page. The gain figures in table 3b look to be off by one stop, ie the gain 4.92 at ISO 50 looks to be about the same as the gain I measured (5.1) at ISO 100. Correspondingly, his read noise figures in table 4b look to be about half of what I see (should be about 4 electrons at high ISO rather than the values of about 2 electrons which are quoted). So I think he made a clerical error early on with the gain and that has propagated through to the rest of the quoted values for the 1D3. Really, read noise in electrons for the 1D3 is not all that different from the 1D2; Canon did quite a good job on that score already in the previous generation; the only thing they improved on was to lower the pattern noise somewhat.
The DR figures Clark quotes at high ISO are OK though, I get about 10.3 stops at ISO 1600 and about 9.3 stops at ISO 3200, about the same as Clark exhibits; I think that means that his clerical error cancels out when taking the ratio of raw saturation to read noise to get DR.
The improvement in DR comes basically from the improvement in gain -- 3.3 electrons/12bit ADU at ISO 400 for the 1D2, 5.1 in the same units for the 1D3 -- 2/3 stop better for the 1D3. Balancing this in the other direction is the change in ISO calibration on the newer Canons which gives back 1/3 stop of this improvement. Finally, it is important to remember that these figures are at the pixel level; if we look on a per area basis, the 1D3 is getting this DR and S/N ratio with pixels that are 36% smaller, and so gets nearly a half stop additional advantage if we compare equal size patches of the frame. Putting it all together, ignoring the change in ISO calibration the 1D3 outperforms the 1D2/1D2N by a bit more than a full stop on a per area basis, perhaps 2/3 stop or a bit more if one accounts for the change in ISO calibration.
These results are tabulated in tables 8 and 9 of
http://theory.uchicago.edu/~ejm/pix/20d/posts/tests/D300_40D_tests/
Note from these tables that the 1D2/1D2N is really a rather poor performer, masking its dismal quantum efficiency by having huge pixels, and thereby doing OK on measures that are pixel-based rather than image percentage based (as they should be IMO).
|
