M8-why 10MB-vs-DMR 20MB

[johnll]

Advertisement (gone after registration)

It's not speculation, it is based on translating the .dng files that have been released using firmware 1.0.6, said to be "production level."

(snip)

Well, if you use 8-bits, you only have 256 levels between black and white (in each channel). By using non-linear mapping, you would be able to put more of the information where you need it most (such as in the shadows or highlights?) but when you convert from RAW to TIF you'll still be using only a small fraction of the 2^16 levels available in each channel.

I am told the M8 DNG files are about 10MB from a 10.3MP sensor. By comparison, RAW files from a Canon 5D which has a 12.7MP sensor range from about 11MP up to over 16MP. This depends on the (lossless) compression which depends on the amount of detail in the image. There are many forms of (lossless) compression and some are more aggressive than others. With no compression one would expect a 10MP sensor to produce files of 20MB at 16 bits. As you surely know, RAW files only contain one channel and the rest of the info comes from "demozaicing" the Bayer matrix. With compression, 10MB from a 10.3MP sensor seems perfectly feasible, even at 16 bits, but in the final analysis, what really matters is how good the prints are.

[Andrej Kolev]

John, regarding "By using non-linear mapping, you would be able to put more of the information where you need it most" - at the out of raw pipeline you usually get 8-bit image, non-linearly coded. This nonlinear power-law is usually named 'gamma' and it is similar to M8 square-root law, altough usually it contains linear part in shadows (Rec. 709). You can change parameters of the gamma curve and emphasize needed tonal range. In the Leica sq.root coding scheme the law is strictly defined, while maybe it gives an estimation with good properties (it seems there is some modifications, some of them are covered by patents). But IMHO, we can get any mapping from original linear data with minimal errors, either sq.root or any other more suitable for our purposes. In case of M8 LUT we are more limited with precision.

 

P.S. John, of cource, what really matters is how goods the prints are. Simply we are also investigate some general properties of nonlinear coding, not only regarding to M8.

[johnll]

Andrej ... yes, I understand the gamma issue (I was just pointing out that one could have variations on that theme), but I still feel uncomfortable with the apparent bit depth of 8 in the DNG files. Why bother to capture at 16 bits (even more than most DSLRs, which mostly capture 12 or so) if you are immediately going to throw away half of them {which represents a huge fraction of the available levels 1 - (2^8)/(2^16) }? ... and if the files are only 8-bit, why are they so large (10MB i.e. uncompressed)? Lossless compression is well known; there are numerous algorithms in common use. And whatever magic you apply, you cannot recover the lost data. Maybe you can minimize its impact, but you can never get it back. We can certainly get "any mapping from the original data" but I'm not convinced the errors are going to be "minimal" under all circumstances. It just does not make sense not to use most, if not all of the information captured by the camera. And Leica have expressly stated somewhere in the specs that the DNG files have 16-bit "resolution" (and if that does not mean 16-bits of data I have no idea what it does mean). I guess we'll find out eventually, but it would be nice to have some clarification from Solms.

[Jamie Roberts]

I mainly take photos for kids and family etc ... but I'll see if I can find a "landscape" somewhere in my archive if you need one, Andrej ... Jamie, I thought you're not too far away from me, our fall colors have gone.

 

Hey Simon--where are you?

 

I've been waiting for the sun, Andre, which seems to have vanished forever here and is now going to hibernate with the coming snow. I did take some shots, though, which do excercise the DR of the camera, pretty much, though they're more muted than I would like.

 

Still, there are enough saturate items to make them useful. They're not what I'd call beautiful by any stretch. But they'll work for the experiment.

 

I'll send them to you over the weekend.

[Andrej Kolev]

OK, Jamie, I'll be waiting.

[gogopix]

well, according to Michael Reichmann, the files are 8-bit

also someone with a dcoder found this to be the case

maybe they are reading the data wrong, the files surely look good!

[Andrej Kolev]

Advertisement (gone after registration)

Victor, decoding is right, you can not get good image without doing right decoding. Image looks good and it really is good, camera, sensor, optics is good. The question is not "why image is so good",but will an image be so good in some complex case after postprocessing. Most likely the difderence will be too small.

[gogopix]

Dear Andrej

 

I think you are right. Most use 16 bit for post processing so there not truncating errors. If leica starts with 14 bit and maps non linear to 8 bit, and then a 16 bit tiff is exported, then PP will play in 16 bit world.

 

THEN, when you LOOK at the image it is usually 8 bit, on screen or in print. So, as you say, the intermediate step of compression to 8 bit (with 14 bit dynamic range,) may never really lose much information.

 

That seems to be the case here. There may be VERY specially cases (extreme DR in ONE color for example, like green leaves only) where someone with loup will find a difference, but so far, the evidence is that these are really fine images!

 

regards

Victor

[johnll]

My guess (and that's all it is) is that the demozaicing algorithm extracts the 8-bit DNG data and computes the color values for each pixel to 16-bit precision. This approach will generate intermediate values between the 256 8-bit steps. I'm not sure I understand "14-bit dynamic range" unless it means 14 stops, which seems to stretch the imagination. What's the connection between brightness levels in the file and dynamic range of the sensor?

[gogopix]

My guess (and that's all it is) is that the demozaicing algorithm extracts the 8-bit DNG data and computes the color values for each pixel to 16-bit precision. This approach will generate intermediate values between the 256 8-bit steps. I'm not sure I understand "14-bit dynamic range" unless it means 14 stops, which seems to stretch the imagination. What's the connection between brightness levels in the file and dynamic range of the sensor?

 

It appear that these kodak sensors can produce values from 2^0 to 2^14 linearly not 2^16 not a big deal. That then is the dynamic range. the algorithm the maps 2^14 to 2^8 and evenything below, but according to a log scale that spreads the brighter data out. This is well know to those who point out that the darker values are lower in number anyway in the low order bits, so it brings the 'detail' of dark and bright areas more into line. It may be this that causes little degradation in image quality, that is, the bits are used where they matter to the eye. So maybe a bunch of values map from the 14 bits to ONE value of the 8 bit, bit your eye may not see that difference anyway!

nothing to do with 14 stops.

[Otto_Fosch]

see this: Phaseone IIQ RAW file format

Phase One -

maybe Leica use the algorithm

 

Frank

[scott kirkpatrick]

It appear that these kodak sensors can produce values from 2^0 to 2^14 linearly not 2^16 not a big deal. That then is the dynamic range. the algorithm the maps 2^14 to 2^8 and evenything below, but according to a log scale that spreads the brighter data out. This is well know to those who point out that the darker values are lower in number anyway in the low order bits, so it brings the 'detail' of dark and bright areas more into line. It may be this that causes little degradation in image quality, that is, the bits are used where they matter to the eye. So maybe a bunch of values map from the 14 bits to ONE value of the 8 bit, bit your eye may not see that difference anyway!

nothing to do with 14 stops.

 

I took a quick look at the spec for the Kodak 10500 chip. Each cell can hold at most 60,000 electrons (that translates into 1.5 volts output), but there are noise sources of about 20 electrons. So while you might want to use 16 bits to count all the electrons, the last 4 bits are just noise, not related to the light falling on the cell except on average. Andrej plotted up the LUT in an M8 dng file, finding it to be just the square root of the original data. Since this is close to the gamma transform used when converting linear light intensities into what a display or printer will produce, it seems this does put the bits where they are needed perceptually.

 

BTW, the chip spec says that there is IR filtration done in the cover glass that comes with the chip. Don't know where the strong red response shown in the LL article is coming from.

 

scott

[scott kirkpatrick]

see this: Phaseone IIQ RAW file format

Phase One -

maybe Leica use the algorithm

 

Frank

 

Nice observation, since Leica has worked closely with Phase One to ensure that good development tools would be available at the time M8 is released. (This is a nice contrast to Canon, Nikon, and Olympus practice, where they try to protect the market for their own software toolkits.) There are two IIQ formats mentioned in the press release, "IIQ Raw Long" using lossless compression, which should produce variable length files, and "IIQ Raw short" which produces fixed length files, said to be one-third the size of the original raw data. Since these are proprietary, patented formats, used in the Phase One digital backs, and DNG is an open source specification, IIQ Raw short is probably not exactly what Leica is doing, but it must have had some influence.

 

scott

[johnll]

So then, if it's 16 bits total (Victor) and 4 bits of noise (Scott) then we have 12 bits of real info, which if the response is linear isn't that equivalent to 12 stops of DR? Very impressive, if true - as good as negs. But we still haven't really figured out how this is obtained from an 8-bit DNG, unless enough extra info comes from the demozaicing.

[scott kirkpatrick]

It's in there. The range of the numbers you get is 12 bits of useful information (that's 4096 values). But only 256 distinct pixel values will be generated when the dng file is read out, spread across this range. Demosaicing will smooth them out, and dithering (adding random low order bits to restore the noise) is another possible trick.

 

scott

[johnll]

see this: Phaseone IIQ RAW file format

Phase One -

maybe Leica use the algorithm

Frank

Very interesting - thanks for the reference. What they say (in the bit that seems possibly most relevant to the present topic) is that the IIQ RAW Small format is "based on" (my quotes) the full 16 bit data captured by the CCD, but is not 100% lossless. They affirm that "in 99% of all jobs" using IIQ RAW Small one will "not be able to see a quality difference".

Well, for that matter, if you get the exposure right and the scene is not too demanding and you don't need much PP, you won't see a quality difference shooting best quality JPEGs either. I'm not suggesting this is all smoke and mirrors - they probably have done some very clever things, but there is no way you can get more than 8 bits of information from 8 bits of data.

Now when you do the demozaicing, you are already converting a single RAW data channel into three color channels, using information essentially hidden in the algorithm. You don't get the same number of bits of information in each color channel as you had in the single channel of data, but it is evident from practise that you get enough to live with.

What I am guessing is that retaining the extra bits that come from doing this at 16-bit precision - even if they are essentially noise - is enough to hide the artefacts that would otherwise become apparent after PP.

In practise, the consensus seems to be that the M8 is producing great IQ, which makes all this discussion pretty well philosophical - however interesting.

[johnll]

(snip) and dithering (adding random low order bits to restore the noise) is another possible trick.

 

scott

Yes, indeed, I hadn't thought of that.

[Andrej Kolev]

Phaseone IIQ RAW - Altough it is celar that in M8 is using DNG non-linear scheme, it may be close (ineresting is the patent description available). I'm suspecting that all that "patented" "new" compression etc. are actually slightly modified methods developed by Soviet and American "space" folks in good old times of cold war.

 

John, regarding get "more than 8 bits of information from 8 bits of data" -it is a question of linear/nonlinear encoding. If we have 12 bits of linear data and if we will leave only 8 of them, we gen a huge loss. But if we will encode it nonlinearly the loss will be minimal, and we will be able to gat good estimation of orginal data.

 

"You don't get the same number of bits of information in each color channel as you had in the single channel of data" - what you mean? I know something about it, so maybe I'll be able to give you some comments, but I dont understand your statement.

 

Scott, " dithering (adding random low order bits to restore the noise) is another possible trick." - dont know, i think the sound and visual perception is different to some extent. I've thought about it but never tried to implement.

[johnll]

(snip)

John, regarding get "more than 8 bits of information from 8 bits of data" -it is a question of linear/nonlinear encoding. If we have 12 bits of linear data and if we will leave only 8 of them, we gen a huge loss. But if we will encode it nonlinearly the loss will be minimal, and we will be able to gat good estimation of orginal data.

 

"You don't get the same number of bits of information in each color channel as you had in the single channel of data" - what you mean? I know something about it, so maybe I'll be able to give you some comments, but I dont understand your statement.

Andrej

Your first paragraph: Yes, if you skew the data in some way that concentrates on the range in which you need more info when encoding, you can certainly enhance the utility of the info you preserve. I was just stating a fundamental fact of information theory. This approach depends on having a more-or-less repeatable pattern of data ... you have to know in advance which range to concentrate on (which, of course, to some degree, we do in photo image data).

Second paragraph: With a RAW file you are just recording the signal at each (filtered) pixel site. The demozaicing algorithm is an attempt to reconstruct what the RGB signals would have been at each site if all three of them had been actually recorded. While results in practise are very good, they are still only an estimate, and therefore strictly cannot contain (say) 3 times 8 bits of RGB information extracted from (say) 8 bits of signal intensity. The demozaicing algorithm is incapable of perfect restoration of the (unrecorded) RGB data. Again, this is just an information-theoretical observation. If the demozaicing computation is carried out to (say) 16-bit precision, you are creating noise, not information, but it is beginning to look like that is enough to smooth out the "contrast jaggies" that can arise when you do a significant amount of PP on (say) a JPG which is 8 bits *only* (and often agressively compressed to boot).

I am sure you know a great deal more about the ins-and-outs of digital images than I do, and I much appreciate your offer ... in the final analysis all this is a case of my wonderment at how much can be achieved with so little real hard data (if you go about it in the right way, of course!). Sorry to be so verbose.

[Advertisement]

Hallo,

Look here: Interesting Leica gear

[Jamie Roberts]

Andrej,

 

DMR raw files for comparison are at yousendit.com, at this address:

 

YouSendIt: The Leader in File Delivery.

 

There are 6 of them, not very interesting as pictures, but I did try to use the whole range of the sensor (and include some saturated colors, which wasn't easy given we haven't see the sun in weeks now!)