brainiac wrote:
I have a different philosophy to expose to the right:
- if contrast is within the DR range, then expose to ensure highlights aren't blown
- if contrast exceeds the DR range, then expose to show the tonal range that you are interested in as you want it to look in the final print
Somewhere between those two methods is usually right.
Exactly how I see it. Furthermore if anyone would like to experiment exposure still affects hue and color saturation. Just like we used to do with 'chrome a bit of underexposure does wonders for the richness of the photo. For reasons that escape me, exposing to the right and reducing the exposure in post does not give the same results. I'd love to hear if anyone else has observed the same thing. Two sides - the technical side and the artistic side. I choose the artistic. I'm too much occupied with what I'm trying to create to worry about all those kibbled bits - had enough of that when I worked in that field.
kidtexas wrote:
The long and short of it is: "Ideally, the noise should slightly exceed the quantization step, in order that roundoff errors introduced by quantization are negligible, and that no bits are wasted in digitizing the noise." In the current crop of cameras (well, up to the 1DSIII and D3), 12 bits essentially does that.
I can see the premise but my take was that he asserts the lower bits are noise and proceeds to do a gradient in photoshop with those bits "removed" to "prove" that it results in posterized levels. The noise added sort of camouflages the steps. I'll have to review the data but I'm not entirely convinced that the sensor is such a bubbling cesspool of photonic noise. I can say that the conversion is for all effective purposes noise-free - modern AD's are pretty dang good. As far as roundoff errors well - that represents one bit - the leastmost bit. Significant only to the range of the conversion and magnitude of the signal. Like I said - I'll need to read a little deeper into the article and really pay attention. It really is an interesting piece.
Oh, it's not that the sensor is a bubbling cesspool of noise. It's light itself. Photon statistics says the photon noise goes like the sqrt of the number of photons.
BubbaJon wrote:
This seems like you're saying that the number 51 doesn't exist because it lies between the fifth and sixth power of 2. It is perfectly reasonable to say that a camera that records 51 levels between black and white captures around 5.6 bits of DR. It is also true that without the ability to process at least 6 bits of data some of the information that this camera would otherwise record is lost because it takes 6 bits to represent the number 51. In terms of processing power there's no such thing as partial bits because transistors (each transistor representing the ability to keep track of another power of two) can't be partially added to a circuit, but the data itself is not restricted to whole values of the powers of 2.
This might be just semantics, but your formulation as quoted above seems confused about the math....Show more →
DMonteith wrote:
This might be just semantics, but your formulation as quoted above seems confused about the math.
In a sense it's semantics but seriously - a bit is on or off. No such thing as .7 of a single bit. Like saying half of an atom or half of an electron. A bit is fundamental. You can make a correlation that DR f-stops correlate to the number of bits but you cannot casually interchange the units. Say you correlate the stock market trend with the price of a barrel of oil - you can't say the stock market is down 200 barrels of oil.
In a sense it's semantics but seriously - a bit is on or off. No such thing as .7 of a single bit. Like saying half of an atom or half of an electron. A bit is fundamental. You can make a correlation that DR f-stops correlate to the number of bits but you cannot casually interchange the units. Say you correlate the stock market trend with the price of a barrel of oil - you can't say the stock market is down 200 barrels of oil.
Fair enough. But this from Wikipedia is interesting:
A bit is a binary digit, taking a value of either 0 or 1. Binary digits are a basic unit of information storage and communication in digital computing and digital information theory...The bit is also a unit of measurement, the information capacity of one binary digit.
That last bit(pardon the pun) is interesting because DXO is obviously using the term as a unit of measurement of information capacity, not as a fundamental element of information storage. Since any unit of measure can be expressed as fractions thereof (inches, meters, seconds, etc.), why not bits? If a 2x4 is 1.5 inches thick then it is not 1 inch or 2 inches thick and if the information capacity of the D3X is 13.7 bits then it is not 13 bits or 14 bits.
Anyway, I'm now sure that it is semantics and not math at issue here and, as a bonus, everyone seems to be right! Now, where's my camera?
DMonteith wrote:
Fair enough. But this from Wikipedia is interesting:
A bit is a binary digit, taking a value of either 0 or 1. Binary digits are a basic unit of information storage and communication in digital computing and digital information theory...The bit is also a unit of measurement, the information capacity of one binary digit.
That last bit(pardon the pun) is interesting because DXO is obviously using the term as a unit of measurement of information capacity, not as a fundamental element of information storage. Since any unit of measure can be expressed as fractions thereof (inches, meters, seconds, etc.), why not bits? If a 2x4 is 1.5 inches thick then it is not 1 inch or 2 inches thick and if the information capacity of the D3X is 13.7 bits then it is not 13 bits or 14 bits.
Anyway, I'm now sure that it is semantics and not math at issue here and, as a bonus, everyone seems to be right! Now, where's my camera?...Show more →
You took it tad too far. It *is* a math issue. Look at the bolded text and repeat after me - bits are basic units. They cannot be made fractional. It even goes so far as to define the unit of measure it's either 0 or 1. There is no in between. Ever hear of integers? If you define something as being in the class of integers then that excludes representations that are non-integer. If that doesn't make sense suggest you look around for a fundamental class in binary.
I've worked with very high-resolution ADCs in the past, this information is not just a repeat of Wikipedia.
In hardware designs, usually an A/D converter delivers LESS than it's full quoted number of bits. Noise, distortion, etc will eat away the Least Significant Bit(s). In datasheets you might see a term called "ENOB", it stands for Effective Number of Bits. This tells you in "real-life" under best case conditions, what your ADC will deliver after you do all the vodoo needed to make the part work.
Typically we measure ENOB by applying a steady input, typically fixed frequency sinusoidal signal of known voltage(amplitude) from a very high precision source. The ADC is run in its typical circuit and you collect data, and perform a Fast Fourier Transform (FFT) on the digital information and analyze the "spectral content" of the circuit.
The FFT x-axis is frequency (in units of Hertz or Radians/sec), and the y-axis is amplitude, usually listed in dB (decibels). FFT is a "digital" version of a Fourier transform, which takes data from the time-domain to one of the frequency domain.
The FFT data will show a huge spike at the input frequency, and potentially other small spikes at multiples/ratios of the input frequency since the ADC is not ideal. ENOB measures the Signal-to-Noise ratio of this peak, to all the other peaks that might show up in the spectral realm. ADCs can have many non-linear processes, so the other spikes might be the result of a complex mathematical interaction between the input and the circuit (i.e. the unwanted spikes could be caused by amplitude/frequency/phase modulation).
The bottom line: The ENOB calculation *CAN* result in a value with a fractional component to it, even though we are talking about a binary world. Since we're performing computations and looking at actual results, we get less than "perfect" numbers. The best 16-bit ADC I measured (in a circuit I worked on) had an ENOB of 15.2 bits, computed using all industry accepted practices.
You can look at these sources for a little more math and confirmation of the "fractional" nature of ENOB:
FYI, if you read those following links they assume you have some background in sampling and quantization theory. Each bit you add to a converter provides you *approximately* 6.02 dB of dynamic range for a given input. See this reference for more information:
Finally, I found this an interesting read on the subject of noise, SNR, ENOB, etc as it directly pertains to Canon/Nikon discussions. I didn't go through all of his results, but the graphs and figures are quite telling:
Note: At the top, I think he meant Noise Power adds linearly. Noise voltage will add with the Root-Mean-Square of the components (square-root of squared sums).
This discussion (about fractional values of DR) is kind of ridiculous... :-) Even if the sensor was made from ONE pixel - the DR could be a fractional value. A digital value with noise added (either electronic or photonic) will have to be averaged over a couple of samples (measurements) if you want a reasonably accurate measurement result. This gives you fractional values as 4 measurements my give a "5" result, 2 each give "4" and "6" results and one give a "3" result. Averaged over the existing nine measurements, this will give you a fractional result ON ONE PIXEL.
And a sensor is NOT made out of ONE pixel, it's made out of quite a lot of them. The average noise of, say, a 200x200 pixel area will of course NOT give an exact unit value, but a fraction valid to at least the second decimal.
Thanks for the links slrl0ver.
One of the differences between the D3x and the 1Ds3 is quite forgotten though, and that's the difference in what value the camera measures as "medium grey" for the exposure. The Canons usually use ~7-7.5% linear value as "medium grey" in autoexposure, and Nikon and some others use ~11%. Sony/Oly mostly use 15-16%. This is most of the difference in the "real ISO" as measured by DxO....
On reason for Canon choosing to do things this way is that the system optimizations that they choose give very poor colour fidelity when you get closer to saturation on the sensor. They NEED more headroom to be able to guarantee colour fidelity in the highlights. As a flipside, they also get less "foot-room" downwards if you choose to let the camera do the exposure, it's already closer to rock-bottom than the other makers. Hence the "pushed shadows" difference between the D3x and the Canons.
And no, you don't ever see this documented if you don't dig really deep into the information, DPP, LR, DxO, CO - they all "compensate" for this difference in "normal exposure" between the camera brands/models without ever telling you.
BubbaJon wrote:
Exactly how I see it. Furthermore if anyone would like to experiment exposure still affects hue and color saturation. Just like we used to do with 'chrome a bit of underexposure does wonders for the richness of the photo. For reasons that escape me, exposing to the right and reducing the exposure in post does not give the same results. I'd love to hear if anyone else has observed the same thing. Two sides - the technical side and the artistic side. I choose the artistic. I'm too much occupied with what I'm trying to create to worry about all those kibbled bits - had enough of that when I worked in that field. ...Show more →
yeah i think it has to do with how the tonal curves in most converters have been designed and you might need to make a somewaht fancy curve yourself to make ane extreme exposed to the right pics pull back down and look completely normal, at least with many converters. Also, I read that many converters use twisted profiles that change the color balance somewhat depending upon signal intensity and extreme exposure to the right and shift things.
kidtexas wrote:
Oh, it's not that the sensor is a bubbling cesspool of noise. It's light itself. Photon statistics says the photon noise goes like the sqrt of the number of photons.
ichiro17 wrote:
DXO Mark is a joke. Its heavily Nikon biased (maybe its true, but I doubt it) and I don't see any proof their tests are sound. But maybe thats just me?
well their results for 40D,50D,5D2 seemed reasonaby similar to what i jsut got for DR myself (they had the 50D,5D2 tail off a little more at high ISO and the 5D2 had a very weird 1.4 stop drop from 12800 to 25600), but generally it seemed like everything was in order
i dont think they did anything weird with nikon SNR/color/etc. the only possible issue might be low ISO DR with some nikons if something nikon did tricked what they did, but there is no evidence showing that this is the case so it might be correct too.
I think this is the in-camera processing the Nikon employs. I have seen tests where the lab was able to compare RAW from D3/D700 with 5D II without any in-camera noise, and they were quite similar.
In my experience, I shoot a D700 and 5D Mk II, and the Nikon RAWs are a little cleaner straight out of camera, but at the expense of sharpness. The Canon RAW are a little noisier but much sharper. A quick NR with DPP and the Canon are quite excellent.
All Canon and Nikon high-end bodies are capable of stellar images. What's curious to me is how little y ou hear about the D3x. The D3/D700 seem to get a lot more buzz/attention than the D3x.
Well, those are my thought's. If the D3x is truly far superior to Canon in DR. then I will look seriously at the D700x at around $3K.
beepclick wrote:
If the D3x is truly far superior to Canon in DR. then I will look seriously at the D700x at around $3K.
If Nikon releases a D700X, they'll be committing financial suicide. Look at what the D700 release did to the sales of D3, and how Nikon is suffering under the current economic climate. Canon also suffers a steep decline in profit, but at least they're still in the black due to strong sales figures from the photographic department.
Anyway, back to the topic at hand, I am not surprised that Nikon has done something to their RAW files. Surely, they'll have taken a leaf from Sony's book of RAW manipulation. However, regardless of how it's done, one cannot help but be impressed with the dynamic range performance of their recent cameras.
If it is indeed true there is some in-camera manipulation of Nikon RAW files, do not be surprised if Canon follows suit. Canon jpeg files were previously known for weak noise reduction and good sharpness, but look at their recent jpeg files. Canon is currently limited by the slow performance of their DIGIC chips as compared to Nikon. But once they fix that, we'll start to to see in-camera Canon RAW manipulation.
Nikon did catch canon napping but lets give credit where credit is due, it wasn't nikon that caught up it was Sony, I give canon credit for designing and building there own chips if Sony hits a snag nikon is screwed.
thw2 wrote:
If Nikon releases a D700X, they'll be committing financial suicide. Look at what the D700 release did to the sales of D3, and how Nikon is suffering under the current economic climate. Canon also suffers a steep decline in profit, but at least they're still in the black due to strong sales figures from the photographic department.
I think nikon suffers much more from marketshare loss in the low end (see for instance the surprising Japan market sales for 2009) than loss of D3 sales after the D700 came out. In fact I am sure with the D700 they probably have been able to get more Nikon shooters going to full frame and made a lot of money from selling more of their excellent but expensive FX zooms. I have many friends that would never buy a D3x, but would jump on a D700x with the same sensor if the price is right (ie less than $3500).
However, regardless of how it's done, one cannot help but be impressed with the dynamic range performance of their recent cameras.