Ad
Ad
Ad
Pages: [1]   Bottom of Page
Print
Author Topic: FujiFilm X-trans Raw Demosaicing - Capture One To The Rescue  (Read 9869 times)
rasterdogs
Jr. Member
**
Offline Offline

Posts: 85


« on: January 02, 2013, 01:30:06 PM »
ReplyReply

The on-going struggle with demosaicing the Raw files from FujiFilm's Xtrans Sensor may be getting some help from Captue One:
http://www.fujirumors.com/capture-one-when-the-going-gets-tough/

I've been using LR4.2 with ok not great results.  Will watch how all of this evolves.

-rasterdogs
Logged
ario
Full Member
***
Offline Offline

Posts: 124


« Reply #1 on: January 02, 2013, 11:13:43 PM »
ReplyReply

There is also a new product, AccuRaw developed by Sandy McGuffog, at the moment under closed beta testing, which looks very promising for the demosaicing of the XTrans Sensor Files (and for other files as well).
Ario
Logged
Vladimirovich
Sr. Member
****
Offline Offline

Posts: 1320


« Reply #2 on: January 02, 2013, 11:40:59 PM »
ReplyReply

basically nobody can decisively do better than SilkyPix which was helped by Fuji from the very beginning = there will be no magical solution to XTrans CFA, not from P1, not from Sandy, not from RPP authors, not from green aliens... so relax and wait until Fuji will come to its senses and drop XTRans CFA for good in 24mp generation of sensors...
Logged
ario
Full Member
***
Offline Offline

Posts: 124


« Reply #3 on: January 03, 2013, 12:40:15 AM »
ReplyReply

basically nobody can decisively do better than SilkyPix which was helped by Fuji from the very beginning = there will be no magical solution to XTrans CFA, not from P1, not from Sandy, not from RPP authors
From what I have see experimenting the available options I have got a different opinion: it is possible to do decisively better then Silkpix.
Ario
Logged
Vladimirovich
Sr. Member
****
Offline Offline

Posts: 1320


« Reply #4 on: January 03, 2013, 01:13:48 AM »
ReplyReply

From what I have see experimenting the available options I have got a different opinion: it is possible to do decisively better then Silkpix.
then you can probably show us, at least Sandy can't in his blog... he is just getting different (sometimes, from Silky) artefacts even w/ a very time consuming hand tuning to provide an illustration in his blog (w/ all due respect to his noble efforts)... like it or not - there is no magical bullet to cure extra R/B sensel separation in XTrans CFA
Logged
sandymc
Sr. Member
****
Offline Offline

Posts: 270


« Reply #5 on: January 03, 2013, 01:42:35 AM »
ReplyReply

David Mcgaughey did a comparison between AccuRaw Beta 5, ACR, RPP, SilkyPix, etc here:

http://www.dmcgaughey.com/2012/12/31/fuji-x-trans-raw-conversion-accuraw-takes-a-swing/

Sandy
Logged
Vladimirovich
Sr. Member
****
Offline Offline

Posts: 1320


« Reply #6 on: January 03, 2013, 02:09:51 AM »
ReplyReply

David Mcgaughey
seems did not use a fine tuning in SilkyPix v5 (not v3 based that Fuji ships), that is what visible, isn't it  Roll Eyes
« Last Edit: January 03, 2013, 02:11:50 AM by Vladimirovich » Logged
ario
Full Member
***
Offline Offline

Posts: 124


« Reply #7 on: January 03, 2013, 07:16:54 AM »
ReplyReply

then you can probably show us, at least Sandy can't in his blog... he is just getting different (sometimes, from Silky) artefacts even w/ a very time consuming hand tuning to provide an illustration in his blog (w/ all due respect to his noble efforts)... like it or not - there is no magical bullet to cure extra R/B sensel separation in XTrans CFA
For the time being I cannot (NDA), but I will.
Ario
Logged
Vladimirovich
Sr. Member
****
Offline Offline

Posts: 1320


« Reply #8 on: January 03, 2013, 08:51:30 AM »
ReplyReply

For the time being I cannot (NDA), but I will.

 Grin , Perpetuum mobile
Logged
Vladimirovich
Sr. Member
****
Offline Offline

Posts: 1320


« Reply #9 on: January 14, 2013, 02:37:47 PM »
ReplyReply

Grin , Perpetuum mobile


so what do we have with C1  Grin ... oh, yes, some people see better than LR/ACR results... better than in camera JPG or SilkyPix ? in their dreams... that is just one more nail into XTrans coffin (no relation to dcraw author)... there will be no demosaicking capable to do anything about 4 "green" sensels clustered together, that's it...
Logged
finaldesign
Newbie
*
Offline Offline

Posts: 3


« Reply #10 on: January 16, 2013, 12:21:12 PM »
ReplyReply

A note from Wikipedia: Bryce Bayer's patent (U.S. Patent No. 3,971,065) in 1976 called the green photosensors luminance-sensitive elements and the red and blue ones chrominance-sensitive elements. He used twice as many green elements as red or blue to mimic the physiology of the human eye. The luminance perception of the human retina uses M and L cone cells combined, during daylight vision, which are most sensitive to green light. [/]

Vladimirovich: ...wait until Fuji will come to its senses and drop XTRans CFA for good in 24mp generation of sensors......ust one more nail into XTrans coffin... there will be no demosaicking capable to do anything about 4 "green" sensels clustered together, that's it...

Vlad, surely you don't want to mislead people into thinking that the Xtrans sensor's "4 green sensels clustered together" ONLY produces problems for the output image, without providing ANY advantages? You curiously omit describing the advantages of having a lot of green sensels clustered together. First of all, the more green sensels on a sensor, the better the probable luminance accuracy and thus perceived resolution of the sensor for most scenes.

Sure, when you get more green pixels, by taking away red and blue pixels, you sacrifice red and blue color quality in favor of green color reproduction quality. BUT, red and blue color components of light contribute MUCH LESS to perceived luminance or brightness of the scene. So it can make perfect sense to "favor" green light pixels over red and blue, because boosting green quality is a "two-fer". When you favor green, you get better green accuracy and, usually, better overall brightness or luminance accuracy-thus-resolution.

Really another way to talk about the advisability of green pixels is to point out that there is nothing "magical" about having exactly twice as many green pixels as red or blue, the way Bayer sensors have 'em. Who said EXACTLY TWICE as many greens as others is "the best" ratio? Maybe there should be even more greens (giving better brightness accuracy) or less greens (giving better red and blue chrominance accuracy). But your contentless criticisms shed no light on the subject.

Speaking as someone who wrote demosaicing software freeware for the common Bayer sensor pattern over a decade ago, I can verify that the software world has been beating against the Bayer sensor array algorithms for at least 10 times longer than anyone's worked on the 2 year old Xtrans sensor layout. And as a Lightroom and RawTherapee user I can also attest that improvements in Bayer demosaicing in the last 10% of its history (i.e. in the last couple of years) has been significant. With those common sense observations, the burden of proof is on you Vlad to give us some evidence beyond blowhard bluster that there is no more progress to be made in Xtrans demosaicing, even though it's a brand new sensor that almost nobody in academia or most of industry has even seriously tackled yet.

You also disqualify yourself as a rational, impartial commenter to imply that the loss of red and blue accuracy in the Xtrans sensor layout doesn't also confer benefits in (a) fighting moire problems and (b) letting Fuji put out a sensor that is not purposefully gimped in sharpness by a blurring anti-moire filter.

Of course for you to argue with my points you will have to tell us that (a) there are no advantages to having a sensor without a blurring/moire filter or (b) as much algorithm research has been done in 2 years of Xtrans' existence as has been done in decades of Bayer demosaicing or (c) there have been no improvements in raw demosaicing algorithms in the last 2 years of commercial software or (d) red and green pixels figure just as much or more in luminance than green pixels.
Logged
Tim Lookingbill
Sr. Member
****
Offline Offline

Posts: 1208



WWW
« Reply #11 on: January 16, 2013, 01:14:26 PM »
ReplyReply

Quote
First of all, the more green sensels on a sensor, the better the probable luminance accuracy and thus perceived resolution of the sensor for most scenes.

That makes a lot of sense. However, the question I always have about this is whether the heavy green bias that would make up a large portion of the luminance accutence that makes up color detail can be and will be proportionally distributed/shared with the red/blue luminance channel detail similar to what happens for blown highlight detail in one channel borrows from the other channels to rebuild what's lost.

I'ld think a much more thought out and sophisticated demosaicing algorithm would have to be worked out in assuring this.

It's similar to channel mixing borrowing luminance detail in one channel to build or restore what's missing in others without introducing an overall bias and artifacts like color errors and posterization of detail.
Logged
Tim Lookingbill
Sr. Member
****
Offline Offline

Posts: 1208



WWW
« Reply #12 on: January 16, 2013, 01:55:54 PM »
ReplyReply

Here's an example of what I'm referring to as color detail in which I've seen several other Raw converters lose by making it look monochrome or the same surrounding color.

Below is a shot I took of an apple pie I baked still in the oven with diffused daylight acting as front fill on the left and the tungsten oven light in the back acting as primary back light. This is edited in ACR 4.6.

Zooming in at 400% on detail I couldn't see at normal full size view revealed this maroon tint to a dark toasted portion of detail that was getting catch light from the diffused window light on the left and its shadow a more yellow/warm hue from the tungsten (see arrows in 400% crop ACR preview below). At default settings this color detail couldn't be seen. I had to bring it out by adding definition in ACR.

Coming from a painting background this kind of color differentiation is highly prized in the fine art painting community because it is the secret sauce so to speak when applying paint mixing techniques in layers that adds depth to a 2D rendered image. Of course the small web sized image can't reveal this but it is useful in an enlarged print around 36"x24" and up.

I wonder what adding extra green pixels to the sensor would do to detail like this if not handled correctly by leaving it up to the various algorithms engineered into other Raw converters? I'm not a computer programmer so I'm not even sure if this even considered when engineering a demosaicing algorithm.
Logged
Vladimirovich
Sr. Member
****
Offline Offline

Posts: 1320


« Reply #13 on: January 16, 2013, 03:14:33 PM »
ReplyReply

First of all, the more green sensels on a sensor, the better the probable luminance accuracy and thus perceived resolution of the sensor for most scenes.
only if they are spaced... in this case you have a cluster of 4 green filters forming a supergreen sensel and I bet w/ any hardware binning to justify... so there is no gain really and only loss... perceived resolution exists only in Fuji fanboys postings... take any 24mp APS-C regular bayer sensor, remove AA filter and say good buy to Fuji.
Logged
finaldesign
Newbie
*
Offline Offline

Posts: 3


« Reply #14 on: January 17, 2013, 01:45:15 PM »
ReplyReply

The difference is a matter of degree with the Fuji sensor. A Fuji red pixel is only touching 3 blue pixels, but it's also touching 5 green pixels. A Bayer red pixel is touching 4 blue pixels and 4 green pixels. Now, you can try to make a huge deal out of the fact that a Fuji red pixel touches more greens and less blues is somewhat not to your liking...

Finaldesign: First of all, the more green sensels on a sensor, the better the probable luminance accuracy and thus perceived resolution of the sensor for most scenes.

Vladimirovich: only if they are spaced...

I guess what you are trying to say is that green pixels only contribute to resolution and luminance accuracy when they are separated from each other by at least one pixel of some other color. I can't make the slightest sense of such an assertion. Why wouldn't a bunch of green pixels clustered together just give you great green luminance accuracy across that cluster? And since green light is the dominant part of perceived luminance, it's not just a green-accurate cluster, it's a probably-luminance-and-resolution accurate cluster.

True, within the confines of any such an all-green cluster, you would have to do more guessing of the red and blue values (of course "guessing other values by looking at the other colors nearby" is the very definition of "demosaicing"). Fortunately even with the Xtrans pixel layout, even though there are more green pixels than in the more common Bayer sensor, each green pixel is "touching" (if only diagonally) at least 2 red and 2 blue pixels. Meaning that even in the 4-pixels "green clusters" that seem to bother you so much, there are still 2 pixels of each other color fairly nearby each green sensel point...To use for "guessing" the red and blue values to assign to those "green points".

Vladimirovich: in this case you have a cluster of 4 green filters forming a supergreen sensel
This sounds like a concern troll, attempting to mislead. There is no way that any sane demosaicing algorithm would "bin" 4 green pixels of a 16 megapixel sensor (that happen to be touching each other) and call them one huge, low-resolution green superpixel. That would be just turning the 16 megapixel sensor into a 4 megapixel sensor. You don't need to look at the green pixels #2 etc that happen to be touching green pixel #1 in order to estimate #1's green value. You only look at neighboring pixels of some other color besides green in order to come up with an estimate for the other color components besides green that you assign to the green pixel point.  Basically you are simply name-calling Fuji engineers as idiots, not unlike your other name-calling in this thread.

Vladimirovich: and I bet w/ any hardware binning to justify... so there is no gain really and only loss...
Even on repeated re-readings I can make no sense of these words. There is no need for "binning" in hardware or software with the Xtrans layout, every single pixel of the Fuji sensor is "touching" at least 2 and sometimes 3 pixels of all the other colors. Meaning that there is some other-color information fairly nearby to every sensel point from which to form a guess as to all the other colors. Which is just like the neighbor-guessing ("demosaicing") you do with any Bayer sensor. The difference is a matter of degree with the Fuji sensor. A Fuji red pixel is only touching 3 blue pixels, but it's also touching 5 green pixels. A Bayer red pixel is touching 4 blue pixels and 4 green pixels.

Now, you can try to make a huge deal out of the fact that a Fuji red pixel touches more greens and less blues is somewhat not to your liking. Because you think red and blue accuracy is more important than green accuracy I guess. But the Fuji engineers are favoring green accuracy over red and blue, which doesn't make them idiots, or doomed to fail, nor justifies any of your other careless, example-free condemnations.

Vladimirovich: perceived resolution exists only in Fuji fanboys postings...
Information-free namecalling will get you everywhere. Sure, everybody who disagrees with your no-facts, misleading criticisms must be "fanboys". And probably we hate cute little children, too. Do you mean to imply that there is no such thing as "perceived resolution"? If you think that's true, then you must also think that deep shadow detail and resolution is perceived just as acutely as highlight detail in a scene. I guess we have to go over the tiresome lesson that nobody but a CIA analyst cares about anything about "the real" resolution of an image, all the rest of us care about is the "perceived" or "apparent" resolution, which is a function of many other things besides the extinction resolution of an optical system. As a simple example, coarse details that are rendered with excellent contrast might give rise to a "sharper-looking" image than a low-contrast image that might have technically smaller, but more dimly rendered, finest details.

Vladimirovich: take any 24mp APS-C regular bayer sensor, remove AA filter and say good buy to Fuji.
Why don't you do that, Vladimirovich, instead of just emptily talking about it? Go to the DPreview studio test comparison page for the Pentax K-5 IIs, and compare it with the Fuji XE-1 at ISO 3200 (I happen to work a lot at ISO 3200).

The Fuji XE-1 ISO 3200 sample is much nicer to look at than the moire-plagued, speckled-looking Pentax K-5 IIs image. Which doesn't tell me that the Xtrans sensor is "overall better or worse". It depends too much on your priorities. The fact that the Xtran camera takes nicer, moire-free ISO 3200 pictures, and nicer-plus-sharper-looking, relatively moire-free ISO 200 photos than the nice K-5 IIs no-moire-filter camera tells me that the huge Fuji Corp's Xtrans sensor is not the product of child-like idiots as you repeatedly imply.

Yes, the K-5 IIs ISO 3200 photos have somewhat more finest detail than the Xtrans photos. The K-5 IIs photos also have somewhat less finest details at ISO 200. But in either case the Xtrans photos simply look better at a glance than the K-5 IIs, a quality which will appeal to a huge number of people. Really at ISO 200 I prefer the look of the Xtrans sample to the mighty Nikon D600's, despite the D600's slightly higher finest detail resolution, simply because the XE-1 has much less moire in the DPreview sample images (see attached).
« Last Edit: January 18, 2013, 08:18:26 AM by finaldesign » Logged
finaldesign
Newbie
*
Offline Offline

Posts: 3


« Reply #15 on: January 18, 2013, 08:15:37 AM »
ReplyReply

FinalDesign: The difference is a matter of degree with the Fuji sensor. A Fuji red pixel is only touching 3 blue pixels, but it's also touching 5 green pixels. A Bayer red pixel is touching 4 blue pixels and 4 green pixels...

tlooknbill: example of what I'm referring to as color detail...Zooming in at 400% ...revealed this maroon tint to a dark toasted portion of detail...and its shadow a more yellow/warm hue from the tungsten...Coming from a painting background this kind of color differentiation is highly prized in the fine art painting community because it is the secret sauce so to speak when applying paint mixing techniques in layers that adds depth to a 2D rendered image....I wonder what adding extra green pixels to the sensor would do to detail like this if not handled correctly

Well in general as you add green pixels and take away red and blue sensors, as both a traditional Bayer sensor does and the Fuji Xtrans does even more, you get more and more resolution and brightness (and gray tone) accuracy, and less and less saturated-color accuracy. So tlooknbill one would think that the common Bayer sensor layout, where every red pixel is touching 4 blues and vice versa, has an inherent advantage in reproducing the color saturation of finest details compared to the Xtrans layout where a red is "only" touching 3 blues etc.

But if perfect-color-detail-everywhere is your top priority, a Foveon-type sensor would be overwhelmingly different and superior in its potential for the accuracy you want. IF you don't care about the Foveon sensor's struggles with high ISO, i.e. with its need for more light. The heartbreaking thing about the Sigma/Foveon setup right now is that, since it needs more light than other technologies to do its job, heck if you are willing to get more light on the sensor somehow there are other ways to improve a given optical system rather than by switching to a less-sensitive (but more color-detail-accurate) Foveon system. You could for example switch to a larger sensor.

So the real comparison to an APS-C-sized Foveon sensor at F2/ISO 200 might be a larger 24x36mm-sized Bayer sensor at F/2.8 ISO 400, in which case the ol' Bayer sensor would give the Foveon camera a real run for its money for a given angle of view and display size.

Have some empathy for your concern about finest detail saturation and accuracy, but am not sure you right now can optimize for that one aspect of performance without introducing other problems that will affect your images far more. For example, have no doubt that the max resolution of a huge Nikon D600 or D800e sensor is better than a little Xtrans unit, but the little Xtrans photos just don't moire so badly and end up looking better at reasonable viewing distances. Also have no doubt that a Foveon/Sigma sensor would be a bit better at some of the job you're interested in. But what about the fact that it just doesn't do high ISO so well yet, and the kind of limited choice of cameras for that sensor?

The proof for me is in the pudding rather than in theory. Because we tend, in differentiating sensors, to be talking about potential performance in finest details that may or may not be relevant to images which are increasingly only viewed on computer screens. Thus rather than just thinking about demosaicing problems, I am won over to the Fuji Xtrans sensor by looking at fairly well-controlled comparisons such as this one from DPreview. A moire-filterless Pentax K-5 IIs versus Xtrans versus Nikon D600 versus filterless-D800e that am attempting to attach. Where the Fuji Xtrans images simply look nicer to me than all the rest, both at ISO 200 and ISO 3200 (the two ISOs in my own work style):
Logged
maddogmurph
Full Member
***
Offline Offline

Posts: 153



WWW
« Reply #16 on: August 06, 2014, 06:30:43 PM »
ReplyReply

Silkypix makes me want to throw myself off a bridge.   Cry
Logged

Maddog - Apologetically critical

Calumet // Tachihara // Schneider 90mm, 210mm // 10" // FUJIFILM X-T1 // 10-24mm, 18-55mm, F55-200mm, 35mm, 60mm // Driod POS // Pentax WG3 // Sony RX100 // Nikon D810 // 24mm 1.4, 50mm 1.4

-AO Member (Amateur's with Opinions)
Pages: [1]   Top of Page
Print
Jump to:  

Ad
Ad
Ad