Re: To RAW or not to RAW?

JPS@xxxxxxx wrote:
>Floyd Davidson <floyd@xxxxxxxxxx> wrote:
>>Jeremy Nixon <jeremy@xxxxxxxxxxx> wrote:
>>>Eugene <nospamthanks@xxxxxxxxxx> wrote:
>>>
>>>> Excuse my ignorance, but how can the JPEG possibly have more precision
>>>> in the shadows than the 12 bit RAW file?
>>>
>>>The 8-bit gamma-corrected JPEG has more precision in the shadows than the
>>>12-bit linear RAW file; that the information does not exist in the RAW
>>>file means the extra precision is not used, not that it doesn't exist.
>>
>>Which also means that *no potential* for it exists.

All of the above is talking about an apple: an 8 bit JPEG image
derived from 12 bit linear data.

>I said clearly earlier that I was talking about the file format, and
>it's potential. Are you one of those people that just can't be wrong?
>You implied a simple mathematical reason why 12-bit linear has more
>dynamic range than an 8-bit JPEG.

You are talking about an orange, not the apple above.

>>>In other words, if the RAW file were, say, 16-bit, it could do better
>>>in the shadows than current 12-bit ones when converted to 8-bit with
>>>gamma correction.
>
>>Actually, there are 12 bit linear formats that have more
>>precision than JPEG.
>
>No; they wouldn't be linear if they did. Linear has poor shadow
>precision.

Nice discussion of oranges as a fruit, but the question was
about apples.

>>Whether any given camera uses one though,
>>is up to the design team. It stands to reason, for example,
>>that if the sensors cannot produce valid data in that range,
>>there is no point in encoding it. The data simply does not
>>exist to begin with. It also does not exist in the 12 bit file,
>>and it does not exist in the 8 bit file either. There is *no*
>>potential for it to exist if the sensor does not produce it.
>
>There is much more data in the sensor at lower ISOs than the 12-bit
>digitization can account for. ISO 100 dynamic range in most DSLRs is
>limited on the shadow end by the limits of 12-bit linear data. Even at
>higher ISOs, the noise is a little noisier than it should be because of
>quantization. Quantization makes the noise values swing more wildly
>than they would if they were left in an analog form, increasing the

Quantization produces "distortion" (as opposed to "noise") which shows
up as something everyone calls noise. The larger the quantization steps
used, the "more wildly" the values swing from the introduced distortion.

>contrast of noise.

It increases the *amount* of the noise, by adding distortion that
didn't previously exist.

>Images look best when the inherent noise in the
>system is rendered precisely; not when it (and the signal) are
>quantized.

I'm not sure what you mean by that.

>Here is an example; both of these images have the exact same
>absolute exposure; taken in manual exposure mode with only the ISO
>setting changed between them. They are of a black Camera strap on a
>black guitar bag, The ISO 1600 shot is pretty much what you could
>expect with 16-bit digitization at ISO 100, except that there is
>probably a tad more readout noise at ISO 1600, so it could be even
>better:
>
>http://www.pbase.com/jps_photo/image/40038800
>
>>Just claiming that there are bits available in the 8 bit format
>>to encode such data means *nothing*. There are also bits
>>available in 12 bit formats too, but there are good reasons they
>>are not used. And it stands to reason that if and when they are
>>useful, the formats will be adjusted to do exactly that.
>
>Neither I nor anyone else here has suggested that you can get better
>shadow level precision than the original RAW data in a JPEG rendered
>from it, yet you are arguing against the strawman, apparently in an
>attempt to look correct.

I responded to what was asked, and the responses that resulted,
all of which were about apples.

*You* keep bringing up oranges.

>Just deal with it, your original reason was
>wrong.

It was dead on right. And it had nothing to do with all this
wonderful information you continue to provide about oranges.

>There is not less Dynamic Range in a typical JPEG because it has
>8 bits; the reason is that the shadows are not converted well because of
>noise reduction, and an s-shaped transfer curve, with clipping of the
>highlights typical. The 8-bit JPEG file format is capable of more
>dynamic range than a 12-bit linear RAW file; that's a fact. If a 12 bit
>file contains more real-world dynamic range than an 8-bit JPEG can
>possibly have, it is not linear.

You have shown a modified 12 bit linear coding that may well
have less dynamic range. But that is not inherent in 12 bit
linear coding. Even with 12 bit linear coding it is possible to
quantize the the entire first f/stop as 1 level and fold it
into the second level just as your 8-bit data did in the graphs
that you posted. That of course provides one more bit to deal
with low light levels. Likewise it is not necessary to clip
everything in the lower 128 levels off.

What you have shown is that one particular 8 bit 2.2 gamma
corrected encoding has more range than one particular abbreviated
12 bit linear encoding. But neither define the two that you
label them as above.

>The fact is, it is possible to render a final low-compression 8-bit JPEG
>with literal 2.2 gamma throughout its range from a RAW file that doesn't
>compromise the shadows or dynamic range.

It is also possible to generate a 12 bit linear encoded image
which you cannot do that with.

--
FloydL. Davidson http://www.apaflo.com/floyd_davidson
Ukpeagvik (Barrow, Alaska) floyd@xxxxxxxxxx
.

Relevant Pages

  • Re: dynamic range and A/D conversion
    ... With a linear conversion, the quantisation step ... JPEG, the percentage steps are more equal across the whole range. ... and that is of course the dynamic range. ... accuracy with which a brightness value in the highest f/stop can ...
    (rec.photo.digital)
  • Re: Real-time sinusoidal parameter estimation?
    ... implemented on an FPGA for a real-time control system. ... For a single sinusoid in noise, you have only one non-linear parameter ... You can use a linear least-squares combined with a line search method. ...
    (comp.dsp)
  • Re: Real-time sinusoidal parameter estimation?
    ... How much of noise? ... You can use a linear least-squares combined with a line search method. ... Since you only have one sinusoid, ...
    (comp.dsp)
  • Re: Real-time sinusoidal parameter estimation?
    ... implemented on an FPGA for a real-time control system. ... For a single sinusoid in noise, you have only one non-linear parameter ... You can use a linear least-squares combined with a line search method. ...
    (comp.dsp)
  • Re: Reverse Linear Regression
    ... > value of resistor and the error in the resistance. ... > linear the data is. ... > it would be nice to just type in the values for the slope, ... and add some noise to it. ...
    (sci.math.symbolic)