scott kirkpatrick Posted October 25, 2006 Share #161 Posted October 25, 2006 Advertisement (gone after registration) oops, accidental duplicate scott Link to post Share on other sites More sharing options...
Advertisement Posted October 25, 2006 Posted October 25, 2006 Hi scott kirkpatrick, Take a look here M8-why 10MB-vs-DMR 20MB . I'm sure you'll find what you were looking for!
Andrej Kolev Posted October 25, 2006 Share #162 Posted October 25, 2006 Scott, here is the links: CASSINI Instrument Catalog Information HALLEY MULTI COLOR CAMERA http://pds-smallbodies.astro.umd.edu/volume/hal_0025/giotto/hmc/hmcguide.doc I've also seen some other mentions, but they only stated that this scheme was used. Havent found any information from Adobe (strange), so the only Adobe information was short Linear.Table tag description. It is Interesting, can we get such files with DNG converter? ------------------------- Andrej Kolev, kolevraw.com Link to post Share on other sites More sharing options...
marknorton Posted October 25, 2006 Share #163 Posted October 25, 2006 Interesting - the Cassini paper talks about converting 12 bit to 8 bit values such as I suggested above and further work I have done this morning suggest the errors are very small if this scheme is used. For those wondering (and who haven't lost the will to live), this doesn't actually require a chip to calculate 10m square roots for each picture. Instead, there is a pre-defined table in the firmware containing the 8 bit values to be used for each 12 (or 14) bit value, and converting from one to the other is a breeze. Not so the lossless data compression where Cassini uses dedicated hardware. It may be Leica adopted this scheme for all sorts of different practical reasons and the more I look at it, the more comfortable I'm becoming with it. Link to post Share on other sites More sharing options...
plasticman Posted October 25, 2006 Share #164 Posted October 25, 2006 For those wondering (and who haven't lost the will to live), this doesn't actually require a chip to calculate 10m square roots for each picture. Instead, there is a pre-defined table in the firmware containing the 8 bit values to be used for each 12 (or 14) bit value, and converting from one to the other is a breeze. So this works as an effective 'look-up' table - as used by programmers when calculations (such as trig functions) would take too long to calculate in real-time? Hmmm - must say i'm still a little nervous about down-sampling at all - but on the other hand, the pictures i'm seeing do seem to be telling their own story... I'm VERY interested that you guys who understand this rocket-science (literally so, it now appears!) follow-up to some conclusion. The whole topic is fascinating. One sidenote - if it turns out that the Leica engineers have been super-smart in developing this system, i'd have been much more reassured if they'd just come straight out and told us about it, instead of this rather bogus 16-bit claim (which is equally false for the entire class of cameras, in actual fact). Link to post Share on other sites More sharing options...
scott kirkpatrick Posted October 25, 2006 Share #165 Posted October 25, 2006 Interesting - the Cassini paper talks about converting 12 bit to 8 bit values such as I suggested above and further work I have done this morning suggest the errors are very small if this scheme is used. For those wondering, this doesn't actually require a chip to calculate 10m square roots for each picture. Instead, there is a pre-defined table in the firmware containing the 8 bit values to be used for each 12 (or 14) bit value, and converting from one to the other is a breeze. Some of the readers of this forum are not computer-facile, and perhaps proud of it, but even they might want to know that in going from a 14 bit (numbers from 0 to 16383) pixel to an 8 bit value the 14 bit value, say 9999, becomes the address in the list of 8-bit numbers which might be saved. In a heartbeat, a tiny computer selects the value at address 9999 and stores it (it is probably 99) in the saved DNG file. Going in the reverse direction, value number 99 in the table stored in the DNG file is presented as the actual value of the pixel, and this is 9801. Close, but not exact, and really simple to do. (I rounded down in this, but the actual method does one more trick, which cuts the maximum possible error in half.) scott Link to post Share on other sites More sharing options...
Andrej Kolev Posted October 25, 2006 Share #166 Posted October 25, 2006 But anyway it is not clear yet, and I'm afraid that science teach us, what there is no such wonders, than you get 14 bit value transform it to 8 bit, and dont pay for that. If I find a time I will try to make such trick with my camera (it has a 14 bit out), and will modify my RAW converter to look on difference. Altough I'm afraid that DR of tiny sensor of camera is too small comparing to M8 KAF. --------------------------------------- Andrej Kolev, kolevraw.com Link to post Share on other sites More sharing options...
scott kirkpatrick Posted October 25, 2006 Share #167 Posted October 25, 2006 Advertisement (gone after registration) Andrej, if you need exerimental material, let's ask Simon to post a nice long-scale DMR raw file, in its 20MB glory, for you do the square root and unsquare root transformation on. scott Link to post Share on other sites More sharing options...
Jamie Roberts Posted October 25, 2006 Share #168 Posted October 25, 2006 Andrej, if you need exerimental material, let's ask Simon to post a nice long-scale DMR raw file, in its 20MB glory, for you do the square root and unsquare root transformation on. scott Hey--I'll provide one too... just let me know what you need. Link to post Share on other sites More sharing options...
Jamie Roberts Posted October 25, 2006 Share #169 Posted October 25, 2006 Guy, I suppose you mean more DR from M8. And difference is not 1.5 stops, as you can see. Also dont forget what at first they get data as 14 bit. Guy--this is where you and I have butted heads in the past. If I understand this correctly, and please someone correct me if I'm wrong, more DR from the sensor doesn't necessarily correlate to more exposure latitude. There are other factors, notably noise, that would affect this. I know we've disagreed on this, but compared with the DMR, the 1ds2, in my experience, has almost 2 stops more exposure latitude than the DMR. So if you preserve the highlights (except for smallest speculars, of course) with your exposure on both cameras, then try to get as much shadow detail as possible, the 1ds2 wins every single time and at every ISO--almost a full two stops difference. This shouldn't be surprising, since the 1ds2 has a usable 3200 ISO, or full two more stops than the DMR! You don't hit the noise floor in shadows on that camera till long after the DMR is a bandy, noisy, mostly unrecoverable mess. BUT--and this is the big but for me--the DMR has more color depth / saturation (which is related to sensor DR). The transitions and saturation levels it is capable of capturing seem an order of magnitude better than the Canon, to my eyes, with the same lenses, and they're much more amenable to editing (though you don't have to edit as much, since they're a bit more compressed in tonal range than the Canons). BTW, all you have to do to see this is intentionally shoot them both straight into a harsh light like the sun in the sky (with a good lens, say the 50 lux) and watch what happens: the Canon CMOS sensor bands like crazy as it tried to accomodate the sky gradient--the DMR just looks like film.. it just "goes" to white. Very, very nice. Ok, it looks like Kodachrome, but still very filmic nevertheless. You only get this smooth transition with the Canon in very, very soft light not too far / falling too quickly out of the sensor range. (BTW--many people blow exposures on the 1 series because they're about a third of a ISO mark hotter than they're marked. I find this constantly checking against a couple of Seikonic incident meters in the studio. The DMR is bang on the money, ISO wise). Does that make sense with what you've seen? I mean, we're both waiting for M8s here, and I love my DMR. But it's Achilles heel is the relatively stringent noise floor and not-so-good high-ISO performance. I can't wait for an M8 at a real 1250...I will be in available light heaven (oh, and a Noctilux would help, too!) Link to post Share on other sites More sharing options...
scott kirkpatrick Posted October 25, 2006 Share #170 Posted October 25, 2006 Hey--I'll provide one too... just let me know what you need. I think a nice landscape with a long tonal scale and delicate tones in the sky would be a tough test. Shot at minimum ISO so that grain is not an issue. The idea is that Andrej could read it as a pure 16 bit raw file, and process it in the full palette as well as the 256-tone palette. Then we can see if anyone can tell the difference. scott Link to post Share on other sites More sharing options...
Jamie Roberts Posted October 25, 2006 Share #171 Posted October 25, 2006 I think a nice landscape with a long tonal scale and delicate tones in the sky would be a tough test. Shot at minimum ISO so that grain is not an issue. The idea is that Andrej could read it as a pure 16 bit raw file, and process it in the full palette as well as the 256-tone palette. Then we can see if anyone can tell the difference. scott Hmmm. Ok, I think I can do that. It's fall here (well, nearly winter) so we may get some great saturated colors too. Since you can't always see things like extreme saturation on the monitor (but they make a difference when printing), maybe Andrej could give us a numeric color delta from the "original" in both palettes? That would very interesting. If the delta is very, very small across the spectrum, then I won't actually worry about too too much... I fear that since most video cards / monitors are 8 bits per pixel anyway, you're going to be hard pressed to see anything till you convert to a color space and properly print (with a RIP). Link to post Share on other sites More sharing options...
scott kirkpatrick Posted October 25, 2006 Share #172 Posted October 25, 2006 Both viewing on a monitor and printing involve pretty much 8-bit dynamic range, and use a gamma correction that is not too different from taking a square root all over again. This means that unless you have done some dramatic stretching of the tone curve, the possible pixel values are now uniformly distributed again. Another reason not to expect much. But with experiments, it never hurts to try things and see if there are some surprises. scott Link to post Share on other sites More sharing options...
sdai Posted October 25, 2006 Share #173 Posted October 25, 2006 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. Link to post Share on other sites More sharing options...
Andrej Kolev Posted October 25, 2006 Share #174 Posted October 25, 2006 Simon, Jamie will wait for your pictures. You can send me a mail to info@kolevraw.com . Here in Moscow is usual autumn , t=10 C, but too rainy, so no DR in all. As for preferable scene it is good to have landscape and also high contrast shot with sharp division of tonal ranges like one half not blown bright sky (clouds are welcome) and other half some object not illiminated so good, because I will change output curve to get details in that halfs. Also i would be very grateful" if you will found image with expressed aliasing patterns (moire). Meanwile I'll make some changes in my program. Diference output is possible. Other processing will have small influence since all my processing pipeline consists only of floating point operations. ---------------------------------------- Andrej Kolev kolevraw.com Link to post Share on other sites More sharing options...
sean_reid Posted October 25, 2006 Share #175 Posted October 25, 2006 Mark, I dont know. We only know that 2^14 is the maximum value of back-transformation LUT. But here is some new info - such type of coding was used in imaging modules of several space missions. Some reasearchers stated that such mapping helps to match the quantization level to the photon noise (I've wrote about photon noise before) and helps to distribute quantization errors uniformly across the dynamic range.Also there is some statements that such transform provided a dynamic range of 14 bits in the 8 bits. Will look for more information, but here is the question, if it so simple why other companies dont use it, why they are bothering with a lossless compression, that you think? Speaking hypothetically only, it's possible that Leica is trying something that has not been done before in digital cameras made for the public. I don't know that that's the case, but it's interesting to consider as a possibility. Cheers, Sean Link to post Share on other sites More sharing options...
Andrej Kolev Posted October 26, 2006 Share #176 Posted October 26, 2006 Maybe Sean. But as I've mentioned before, nonlinear coding is proposed by Adobe in DNG, so maybe Leica is simply a first company which has been interested in applying that scheme. It seems that it was widely used by NASA and other goverment structures to save bitrate in space/orbital missions. And for example Cassini team had kept ability to save full linear data. ----------------------------------- Andrej Kolev, kolevraw.com Link to post Share on other sites More sharing options...
johnll Posted October 26, 2006 Share #177 Posted October 26, 2006 I find all this very interesting, but seems to be essentially speculation. I also find it a little disturbing that Leica cannot be a bit more transparent with technical information. However, I still cannot imagine that the M8 records only a bit depth of 8/channel, even if it is non-linear. BTW, there is no relationship between bit depth and dynamic range. The zero value represents black, the maximum value represents the strongest signal the sensor can support, which will map to white. In between you can go linear or not. The bit depth just sets the number of intermediate steps between black and white, and that defines the size of each step, according to the linear or non-linear rules defined for the system. For certain digital manipulations, such as curves, color balancing etc you need finer steps (than 8-bit) to avoid posterization when you return to 8 bits for printing. Link to post Share on other sites More sharing options...
scott kirkpatrick Posted October 26, 2006 Share #178 Posted October 26, 2006 Scott, here is the links:CASSINI Instrument Catalog Information HALLEY MULTI COLOR CAMERA http://pds-smallbodies.astro.umd.edu/volume/hal_0025/giotto/hmc/hmcguide.doc I've also seen some other mentions, but they only stated that this scheme was used. Havent found any information from Adobe (strange), so the only Adobe information was short Linear.Table tag description. It is Interesting, can we get such files with DNG converter? ------------------------- Andrej Kolev, kolevraw.com I've read these and some related papers, and just as you say, they seem to regard the square root transform reducing 12 or 14 bits to 8 bits as a standard practice. I'll have to look in textbooks to see what its presumed strengths and weaknesses are, as there are no references to literature on the technique. Google also comes up empty. scott Link to post Share on other sites More sharing options...
scott kirkpatrick Posted October 26, 2006 Share #179 Posted October 26, 2006 I find all this very interesting, but seems to be essentially speculation. I also find it a little disturbing that Leica cannot be a bit more transparent with technical information. However, I still cannot imagine that the M8 records only a bit depth of 8/channel, even if it is non-linear. BTW, there is no relationship between bit depth and dynamic range. The zero value represents black, the maximum value represents the strongest signal the sensor can support, which will map to white. In between you can go linear or not. The bit depth just sets the number of intermediate steps between black and white, and that defines the size of each step, according to the linear or non-linear rules defined for the system. For certain digital manipulations, such as curves, color balancing etc you need finer steps (than 8-bit) to avoid posterization when you return to 8 bits for printing. 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." I agree with you that the M8 has a 14 or more bit dynamic range, but it does seem to have only 8 bit resolution. scott Link to post Share on other sites More sharing options...
Andrej Kolev Posted October 26, 2006 Share #180 Posted October 26, 2006 Scott, I have an additional info. From FAQs There are several forms of image compression that can be used. One of them is "square-root" encoding (also known as companding) that can be used to convert the 12-bit linear output of the cameras to an 8-bit non-linear output with information loss that is negligible for most purposes. New word companding quickly can help us to avoid "google noise" and we get docs like this: Optimized Shot Noise Companding for MSI ------------------------------------- Andrej Kolev , kolevraw.com Link to post Share on other sites More sharing options...
Recommended Posts
Archived
This topic is now archived and is closed to further replies.