I understand what you're saying (I think) but I still have a problem
with the notion that Canon can apply their "inside information" since
that information has to be encoded somewhere within the raw file. If we
have 14 bits to encode brightness information and underexpose by 4 stops
it seems to me that the lowest 4 stops of the original scene have just
become 0. Those lowest 4 stops have just all been melded into the
blackest of blacks and there is no way to recover them.
If there is a difference (as there is) between ACR and DPP, whatever is
being recovered by DPP has to come from what would have been (in a
normal exposure) some 4 stops higher up the brightness chain...
otherwise it would have been melded into 0. The fact that DPP can
recover something means it was not 0. Since ACR reports 0 for some
values I conclude ACR has missed the boat somehow.
The only way I can see DPP using some inside info here is that the raw
file contains something more than just 14 bit brightness values for each
sensor site. So, where am I missing the boat?
I'll be convinced if the 5D under ACR behaves the same as the 5DMkII
Chuck Norcutt
Moose wrote:
> Chuck Norcutt wrote:
>> I could see that it was wrong without pushing since what should have been
>> varying dark gray values were almost all zero immediately after opening the
>> file in ACR. There is clearly an ACR bug and it occurs immediately on
>> opening the file... no alterations required. The question is whether this
>> is limited to the 5DMkII.
>>
>
> I've been reading, testing and thinking. I think I understand what's
> happening. whether I can explain it clearly or not ...
>
> First, take a look at the two histograms under the heading "Linear
> Histogram", about 30% down this page.
> <http://www.guillermoluijk.com/tutorial/dcraw/index_en.htm>
>
> Fig. 5 shows the linear data from a sensor. F-stops involve
> doubling/halving the amount of light for each full stop, the legend on
> the bottom shows how that plays out. Let me put it in numbers, adjusted
> from the 12-stop range from a 12 bit sensor to the 14 bits of the 5DII.
> The top stop has a range of 8,182 values in which to store levels of
> brightness and:
>
> 14 8192
> 13 4096
> 12 2048
> 11 1024
> 10 512
> 9 256
> 8 128
> 7 64
> 6 32
> 5 16
> 4 8
> 3 4
> 2 2
> 1 2
>
> So - When the subject is underexposed by four stops, the number of
> possible values goes from 16,384 down to 2,048. Not so bad, 8 bit is
> only 256, right?
>
> Ah, but look at the shadows. For a correctly exposed shot, there are 32
> values to hold all the data for the bottom 5 stops. No matter that it
> looks like more after gamma adjustment, it's still those few values
> up-sampled.
>
> In the 4 stop underexposed version, data from some of the higher stops
> have to be squeezed down into those 32 values. So, potential data detail
> is lost, compared to proper exposure. This is true even if the sensor
> can perfectly discriminate the actual values way down at such low levels
> - the data is lost to the digitization process.
>
> To the extent that the sensor loses effective precision in the dark,
> things get worse. Think of a digital meter. It's not uncommon to have
> finer resolution than precision. The meter reads to 0.01, but the
> accuracy is ±0.03. so a reading of 1.35. although it looks accurate on
> the LCD, is really somewhere between 1.32 and 1.38. When you get down to
> very low readings, the precision means values may be 100% or more off
> the indicated value.
>
> I assume the same holds true for values out of the A/D converters in
> cameras.
>
> A shot that is intentionally heavily underexposed has its true shadows
> recorded in only a handful of values and its lower mid-tones compressed
> into maybe 100 values. Then in the RAW converter, we try to pull those
> values back up the where they would be in a proper exposure. That means
> lots of big up interpolation.
>
> To the extent that the sensor system is differentially non-linear
> between channels at very low light levels, the reconstituted image will
> have inaccurate color. To the extent that quantization, what might also
> be called pixelization or rounding effects as analog values were
> digitized down at the bottom, there will be color/luminance irregularities.
>
> Why then, does Canon's DPP do a better job in this test than ACR? I
> think the answer is simple. First, you should read on down in the page
> linked above, on through the "Sensor Saturation Level" section. It may
> make your head hurt, but will give an idea why the maker of a sensor
> system, with full in-depth knowledge of its characteristics, can
> properly fine tune conversion settings for it.
>
> Adobe, on the other hand has to set it's default conversion behavior to
> accommodate a vast array of sensor systems. Sure, they can tune settings
> defaults for each camera based on some test shots, but they can't really
> know what's going on down in the depths of each system.
>
> I tried converting CH's sample color chart image in DPP and ACR. I then
> tried the white balance dropper on various gray steps. In both cases,
> there was a difference between the bright and darker patches. the
> lighter the patch, the more accurate looking the color.
>
> Overall, as expected from the above, the DPP conversion was better than
> the ACR, and the difference became greater the darker you went. The
> darker ACR patches had a distinct magenta cast, Which makes me think it
> may have something to do with the effect on relative sensitivity at very
> low light levels of the use of 2 green sensors for each single red and
> blue sensor - or not.
>
> My conclusions?
>
> 1. It's foolish to expect recovered shadows to be as color and luminance
> accurate as if they were mid-tones. The technology just doesn't support
> it. HDR image file formats use floating point decimal numbers. If future
> cameras adopt a floating point RAW format, this would be different. For
> now, we have to develop shooting techniques to accommodate these
> limitations of our cameras to our photographic purposes.
>
> 2. There probably isn't a programming error in ACR. It just doesn't know
> as much about particular sensor systems as do the manufacturers.
>
> 3. I may try using DPP for images where I care about color accuracy and
> must pull shadows up a lot. On the other hand, the default color results
> I get from ACR on normal 5D RAW files look more color accurate to me
> than the defaults from DPP and ACR gives me more and more useful
> controls. So I'll stick with ACR for most work.
>
> Remember, most of my shooting is of natural subjects, where there is no
> absolute measure of color accuracy. In many cases, the shadows in the
> subject are dark enough that my own eyes are running mostly on B&W
> cones, so even if I had perfect color memory, I might not know what
> color those shadows 'really' were.
>
> 4. While all this is interesting to me in a theoretical way and I've
> learned a lot that may help inform my image creation and processing, I
> don't see how it has had a practical effect on the vast majority of the
> images I've processed. Even where it may have, I'm not sure how I could
> tell.
>
> So I take a sunset shot with underexposed foreground, pull it up and get
> some magenta shift. There's a lot of magenta in the shadow areas of
> sunset scenes as I see them with my eyes. How can I tell the difference?
> Why should I care? If it looks right to eyes that have seen many
> thousands of sunsets, why worry? It it looks too magenta, I can just
> pull magenta saturation down a bit. There really is no absolute in most
> photographs.
>
> Moose
--
_________________________________________________________________
Options: http://lists.thomasclausen.net/mailman/listinfo/olympus
Archives: http://lists.thomasclausen.net/mailman/private/olympus/
Themed Olympus Photo Exhibition: http://www.tope.nl/
|