NEWS: HD Photo to become JPEG XR

Barry Pearson schrieb:

I have published my own tests at the page below. My conclusion from
these (admittedly limited) tests was:

"For any given set of quality values, the HD Photo and JPEG 2000 files
were about the same size, and significantly smaller than the JPEG
file". ("Half the size" would be a fair generalisation, and at that
level probably fewer nasty artefacts). The basis for these conclusions
is at the following page - I won't repeat them he
http://www.barrypearson.co.uk/articl...ysis_lossy.htm


How do you define "quality"? What is a "quality value"? You have to be
*very* careful defining terms, or you're likely comparing apples to
oranges. Sorry to be picky, but one has to be pretty careful when
testing.

Some typical mistakes (not saying that you did it, just things to watch out for):

The "quality" value of HDPhoto has nothing to do with the "quality" value of
JPEG. JPEG2000 doesn't even define something like that, but lets you choose the
target file size. Vendors might prefer to offer a "quality" setting, but that's
then entirely vendor defined.

Others measure "quality" in terms of PSNR (peak signal to noise ratio), then
often logarithmic in dB. This "quality" relates pretty badly to "visual impression",
it is easy to generate two images with the same dB as "quality", one of them
looking great, the other looking unacceptable. Better metrics exist, none of them
is perfect, but most of them are notably better than PSNR.

This said, our tests do the following:

i) Pick a test image set, for all images in the set:
ii) compress the image to quality level x of HDPhoto
iii) measure the file size returned by HDPhoto
iv) compress the same image to the same file size with JPEG2000 (yes, JPEG2000 can
do that)
v) Use either an objective visual quality metric (we've used M-SSIM, VDP and PQS)
or run subjective tests (with people actually looking at the images), and by that
define a quality.

vi) compile results by plotting "perceived" or "objectively measured" quality over
output rate.

When doing this test (and not measuring PSNR and not adjusting obscure quality
settings), you really compare apples to apples. Unfortunately, and that's what I
said, HDPhoto doesn't at all perform very well. And, as said, I hope that this is
fixable, but the current version is not.

Furthermore, please also note that the *current* test image set is still limited,
more tests are to be done.

The major drawback is its lack of standardization. The problem seems that camera
vendors prefer to bind their customers instead of making images interchangeable.
I don't see why JPEG-XR would change their position, but that's all my guesswork.
Otherwise, it would have been easy just to approach the JPEG to standardize *some*
type of raw format - it is IMHO just not desirable for the vendors.

There IS an ISO standard raw file format. ISO 12234-2 (TIFF/EP). (It
became an ISO standard in 2001, and some manufacturers such as Canon
and Nikon based their own raw file formats on it). The problem is that
it was never really fit for the purpose of standardised interchange,
and it has become out-of-date. (In effect, DNG, also based on that
standard, is ISO 12234-2 brought up-to-date and made fit for purpose).

ISO are revising ISO 12234-2, and Adobe have given them permission to
use the features of DNG in the revision. (Just as they gave ISO
permission to use TIFF in the original version of ISO 12234-2).

Given that ISO's TC42 WG18 has responsibility for TIFF/EP, it would be
diversionary for another working group to "compete" to standardise a
raw file format. Far better, surely, for all standards bodies and
working groups to concentrate on one standard, especially if it can
based on a format (DNG) that is supported in some way by nearly 200
products of various kinds. (I would like to see the revised ISO
12234-2 either BE DNG, or be compatible with DNG sufficiently for
products to work with a common subset).

Thanks for the pointers to TC42 (I'm SC29, so there you go), I'll ask
there. In my impression, TC42 doesn't really have too much support
or acceptance from the camera vendors, but I'm (currently) only watching
this from the outside, so I cannot really say for sure. I'd be personally
happy if any type of "raw format" standardization would arrive at a format
that is really accepted and that makes image data interchangeable, but
probably there are too many sensor types, or too many trade secrets - I don't
know.

So long,
Thomas

On Nov 30, 7:48 am, Thomas Richter wrote:
Barry Pearson schrieb:

I have published my own tests at the page below. My conclusion from
these (admittedly limited) tests was:

"For any given set of quality values, the HD Photo and JPEG 2000 files
were about the same size, and significantly smaller than the JPEG
file". ("Half the size" would be a fair generalisation, and at that
level probably fewer nasty artefacts). The basis for these conclusions
is at the following page - I won't repeat them he
http://www.barrypearson.co.uk/articl...ysis_lossy.htm

How do you define "quality"? What is a "quality value"? You have to be
*very* careful defining terms, or you're likely comparing apples to
oranges. Sorry to be picky, but one has to be pretty careful when
testing.

The term "quality values", as I used it, is defined on the above page.
As I said, I won't repeat it here. Specifically, at:
http://www.barrypearson.co.uk/articl...sy.htm#results

[snip]
Given that ISO's TC42 WG18 has responsibility for TIFF/EP, it would be
diversionary for another working group to "compete" to standardise a
raw file format. Far better, surely, for all standards bodies and
working groups to concentrate on one standard, especially if it can
based on a format (DNG) that is supported in some way by nearly 200
products of various kinds. (I would like to see the revised ISO
12234-2 either BE DNG, or be compatible with DNG sufficiently for
products to work with a common subset).

Thanks for the pointers to TC42 (I'm SC29, so there you go), I'll ask
there. In my impression, TC42 doesn't really have too much support
or acceptance from the camera vendors, but I'm (currently) only watching
this from the outside, so I cannot really say for sure. I'd be personally
happy if any type of "raw format" standardization would arrive at a format
that is really accepted and that makes image data interchangeable, but
probably there are too many sensor types, or too many trade secrets - I don't
know.

Some camera vendors (eg. Canon, Nikon) based their raw file formats on
ISO 12234-2 from TC42. Indeed, NEF files identify the TIFF/EP version
used.

There are actually not many sensor types - they are mostly of few
general types (eg. Bayer, Fujifilm SuperCCD) but with many parametric
differences that can be caterd for. That is how DNG works, as shown
at:
http://www.barrypearson.co.uk/articl...n.htm#examples

One "trade secret" problem is with Foveon/Sigma X3F files, where
documenting their details would identify internal details of the
Foveon software used by some raw converters. So DNG only supports X3F
images after using some Foveon software to process it, and doesn't
hold its raw state.

Camera manufacturers COULD use a common standard if they chose, (with
a possible exception for Foveon), which is why a number of niche and
minority manufacturers support DNG. (My Pentax K10D gives an option
whether to use PEF or DNG, and gives the same results either way).
http://www.barrypearson.co.uk/articl...#manufacturers

--
Barry Pearson
http://www.barrypearson.co.uk/photography/

Barry Pearson wrote:
On Nov 30, 7:48 am, Thomas Richter wrote:
Barry Pearson schrieb:

I have published my own tests at the page below. My conclusion from
these (admittedly limited) tests was:
"For any given set of quality values, the HD Photo and JPEG 2000 files
were about the same size, and significantly smaller than the JPEG
file". ("Half the size" would be a fair generalisation, and at that
level probably fewer nasty artefacts). The basis for these conclusions
is at the following page - I won't repeat them he
http://www.barrypearson.co.uk/articl...ysis_lossy.htm
How do you define "quality"? What is a "quality value"? You have to be
*very* careful defining terms, or you're likely comparing apples to
oranges. Sorry to be picky, but one has to be pretty careful when
testing.

The term "quality values", as I used it, is defined on the above page.
As I said, I won't repeat it here. Specifically, at:
http://www.barrypearson.co.uk/articl...sy.htm#results

Well, your quality settings seem to be very low - unrealistically low.
You do not use those rates in real images (or, I actually wouldn't want
to (- The algorithms might and will behave differently for higher
quality. JPEG2000 doesn't have a "quality" at all, it's a vendor defined
thing. May I ask which vendor you picked? Without that, a "quality"
makes no sense at all.

How do you define "examine the differences"? This is not well-defined,
what do I do with these histograms to get a "quality"? Do you
want to limit the maximum error? (Mathematically, this is the l^infinity
metric). Not really a very usable metric - a single wrong pixel can
ruin it, even though if that is in a suitable image part, you wouldn't
even see it. I've here images where the absolute error for JPEG is
pretty huge, but the image is pretty fine. It really depends *where*
this error is in the picture. If you look at the "mean" error, (guessing
that this is the *mean square error*) then that's basically PSNR (which
I already mentioned). It is a fairly bad metric, but a popular one with
known deficiencies.

May I suggest a different way of handling it? Measuring the *file size*
is much easier. Try to adjust compression parameters such that for
all three methods, you get approximately the same size. Then inspect
images visually or by objective metrics. I can send you code for that.

Defining the quality from the histogram is unfortunately not very
reasonable as well - you neglect a lot, if not all visual effects known.
Especially at bitrates *that* low, a lot of nasty things happen.

Examples: Human observers are more sensitive to specific spatial image
frequencies than others (that's also the reason why the JPEG-1 default
quantization matrix looks so weird), are less sensitive to color than to
luminance, and are less sensitive to errors when those errors are in
"noisy" image regions, i.e. errors can be easily hidden behind image
structures.

This is also the reason why every sensible image metric requires
careful "subjective" testing, i.e. ask a lot of people on their
opinion of the quality, then check how sensible the metric can predict
those numbers.

Some camera vendors (eg. Canon, Nikon) based their raw file formats on
ISO 12234-2 from TC42. Indeed, NEF files identify the TIFF/EP version
used.

There are actually not many sensor types - they are mostly of few
general types (eg. Bayer, Fujifilm SuperCCD) but with many parametric
differences that can be caterd for. That is how DNG works, as shown
at:
http://www.barrypearson.co.uk/articl...n.htm#examples

Thanks for that, I haven't really looked into DNG, though I was
aware of its existence (but only that and barely more).

One "trade secret" problem is with Foveon/Sigma X3F files, where
documenting their details would identify internal details of the
Foveon software used by some raw converters. So DNG only supports X3F
images after using some Foveon software to process it, and doesn't
hold its raw state.

Camera manufacturers COULD use a common standard if they chose, (with
a possible exception for Foveon), which is why a number of niche and
minority manufacturers support DNG. (My Pentax K10D gives an option
whether to use PEF or DNG, and gives the same results either way).
http://www.barrypearson.co.uk/articl...#manufacturers

Thus, asking frankly, why don't they? (-:

So long,
Thomas