Reading mode: Light Dark

The truth about 14-bit

Go to last post
Hardnose said:
I often get shadow posterization, pixelation or dense speckled black
out of place patches in the darkest areas of my shots. Perhaps
they're under lit and i'm pulling them up in levels too much.

Would 14 bit on the 40D or 1Dmk3 help this? sounds like no from your
discussions so far.
Click to expand...
That's probably due to the converter or its settings (if you're using ACR, for example, you need to lower the shadows slider to zero if you want all of the RAW shadows). If you're shooting JPEG in the camera, shadows are usually quite posterized, but you can improve the shaowds by using the lowest contrast setting, as that pulls up the shadows and pulls in some extra highlights.

--
John

 
So... then what are the chances that Canon would ever agree with your conclusion and/or would allow the user to opt out of the extra 2 bits of data?

(OK, so I believe I know the answer to that and/or that my odds are better of being hit by lightning...)

But anyway, wouldn't it be nice if the user could select the bit depth saved rather then having this silly sRAW stuff... ? OK, so I know the answer to that too...

but gosh... How can we all just stand by and let Canon fill our memory cards and hard drives with noise?
John Sheehy said:
The 14-bit ADC may or may not have a benefit (max signal divided by
read nosie is the same for both the 1Dmk2 and mk3 at ISO 100), but I
don't think we need to have mandatory 14-bit output, because it
increases the file size tremendously, with 2 bits of uncompressible
noise.
Click to expand...
 
That just shows that there is more than 12 bits of information in the analog data. It actually reinforces the notion that after the A/D converter there are no more than 12 useful bits. Otherwise, you'd just convert with no amplification and just use the lower bits for high ISO shooting.
--
http://www.pbase.com/victorengel/

 
Any numbers that can be derived from this or other images. Maybe it's my eyes, or wishful thinking, but the 14-bit seems to look better to me.
 
It would be more useful than sRaw. Speaking of sRaw, I think it's possible for it to have useful additional bits. It all depends on how it's implemented.
--
http://www.pbase.com/victorengel/

 
ejmartin said:
Daveed V. said:
A good introduction to some of the quantization issue can be found here:

http://www.clarkvision.com/imagedetail/digital.sensor.performance.summary/index.html

As you can see from one of the graphs in there, even the 20D/30D
sensor can in theory benefit from a 14-bit A/D converter, if
amplifier noise is sufficiently controlled.
Click to expand...
A contrary result based on actual 1D3 images is to be found at

http://www.openphotographyforums.com/forums/showpost.php?p=31694&postcount=31

The surrounding thread is also rather illuminating.
Click to expand...
(Coincidentally, I read that thread a few days ago (via someone else's suggestion), but I don't see how it relates.) AFAICT it's a demonstration that Canon "cheats" in its implementation of fractional ISO settings by using digital scaling instead of regulating the local amplifiers.

(Well. the 14-bit A/D convertor may play a role in this in that when you get close to unity gain, the amplifier becomes ineffective for noise control and digital scaling is the best option (and there having the extra bits pays off). That's why the graph in the article you linked to shows monotone behavior in the high ISOs; Canon's unity gain ISO point is traditionally slightly over 1000.)

--
Daveed
http://vandevoorde.com
 
And the first sentence was supposed to be a question.

Where's my caffeine.....
 
I was wondering so I made some calculus.

The 12-bit/14-bit files are light linear intensity, not something gamma corrected.

So let's say roughly that we convert linear intensity to the image by applying a gamma inverse of 2.2, this makes something like
y = x ^ 0.454545

(as far as I remember, this is the light mapping which is used for the DIGIC to make the jpeg files)

where x is the light intensity on input, and y the output (would be the value to read in the info panel in photoshop). x is normalized, that is x goes from 0 to 1, same for y.
This function has the fastest start at the origin.

So now I take a value of 1 in light linear intensity for a 12 bit file to a 8 bit output :
((1 / 4096)) ^ 0.454545) * 256 = 5.84

This means that without changing the exposure, the darkest black is almost 6 in a 8 bit file (except 0)

The same for 14 bits :
((1 / 16384)) ^ 0.454545) * 256 = 3.11

Does not look enough for a gamma corrected 8 bit file in all cases.

To have an enough bit in linear light to have one step in the darkest region of an output 8-bit image would require 18 bit files (!)

Raphael
 
Current consumer grade display hardware (graphics card, monitor) are only 8bits per color, and jpeg files are also 8 bits.

The benefits are by large in the intermediate processing, advanced software can pick the which of the 8bits to use, or it can compute 8bits based on 14 raw data bits data based on some kind of algorithm.

Eventually, I think there will be revolution in monitors/graphics card, so that we can get the 14 bits on the display without trimming the 6 bits of the data.
 
--

lenses make the image, the camera only records it. my zoom is my feet, not my fingers, don't be lazy, buy primes :)
 
John Sheehy said:
Daveed V. said:
A good introduction to some of the quantization issue can be found here:
Click to expand...
http://www.clarkvision.com/imagedetail/digital.sensor.performance.summary/index.html

I don't consider Roger to be an authority on these issues. His role
is to reinforce the myths of yesteryear, IMO.

Just because someone details their experiments and calculations,
doesn't mean that their conclusions are useful or accurate.
Click to expand...
What of his analysis do you disagree with, and why?

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/
 
I can definitely see a difference between the last two. I suggest, though, that the subject is not the best for this sort of analysis. Can you do the same thing with a smooth subject -- one that doesn't have fine detail that would make differences hard to discern? How about a shot of a Gretag Macbeth chart, for example?
--
http://www.pbase.com/victorengel/

 
The more information you can capture is always useful, up to the point you can't manage the data anyway. This should give some headroom in the all important dynamic range (raw files only of course).

Note that this is not the first 14 bit camera, and it has proven well elsewhere.

sunflowerflyer
 
digitalrock said:
John,

Is your example truncation without rounding? Just wondering if there
would be any visible improvement with 1/2LSB rounding prior to
truncation?

Not a criticism, just curious! :-)
Click to expand...
The ideal way to quantize to 1/4 the values is to add 2, divide by 4, multiply by 4, and then subtract 2. This keeps values rounded (as Photoshop's Levels tool does if you use it to effect quantization).

I don't remember if I did that here, but if I didn't, it would make the quantized images worse.

--
John

 
You must log in or register to reply here.

Keyboard shortcuts

f
Forum
m
My threads
Mobile site
About
Editorial content
Cameras & Lenses
Community
All content, design, and layout are Copyright © 1998–2025 Digital Photography Review All Rights Reserved.
Reproduction in whole or part in any form or medium without specific written permission is prohibited.
When you use DPReview links to buy products, the site may earn a commission.
©GPS Media - Guides, Products, Services.
Back
Top