philmoz said:
There's obviously a lot of very knowledgeable people here so I'll pose a question that's been bugging me for a long time.
How is it possible for DxO to claim > 14 stops of dynamic range for cameras with a 14 bit ADC ???
Phil.
Dynamic range has two key interpretations, and they are often conflated resulting in the kind of confusion you see here.
Mathematically, dynamic range is simply defined as the ratio between the smallest and largest (or darkest and brightest) values of some measurable quantity
. In the case of photography, it is more specifically defined as the ratio between the RMS of read noise (read noise specifically, that's important) and the maximum saturation point of a pixel. The reason the root mean square (or some similar average) of read noise is used as the "darkest" value is because read noise intrudes into and eats away at useful values below a certain level. When read noise is higher, the ratio decreases, thus reducing dynamic range.
Another key interpretation is the
bit error ratio (BER). This has to do with noise in the signal at every measurable point between the extremes as defined above, not just in the shadows. When you hear someone say "a full-frame sensor has more dynamic range than an APS-C sensor", they are really referring to a lower BER. A full frame sensor like the 5D III has a full well capacity of ~68ke-, where as an APS-C sensor may have a full well capacity of 22ke-. The 5D III has three times the maximum signal power, meaning it has less noise at all levels, not just read noise but intrinsic signal noise as well. When DXO says an image gains dynamic range by downsampling, it is because the BER was reduced thanks to the averaging involved in downsampling. Not just in the deep shadows, but at all levels of the signal.
These two interpretations are basically two sides of the same coin, but there is a key difference between them that really matters: One has to do with the quality of the signal between the extremes, the other has to do with what the extremes are. Screen DR deals with what the extremes are. Print DR deals with the quality of the signal between the extremes. Read noise has a direct impact on EDITING LATITUDE, and as far as photographers are concerned, that is synonymous with shadow and highlight recovery (really, shadow lifting.) Less read noise, regardless of the maximum signal strength or the amount of intrinsic noise in the signal, the more you can lift shadows without having them
look ugly because of read noise.
Screen DR tells you how many stops of dynamic range you have to work with when editing RAW. Print DR tells you how nicely the signal cleans up when you downsample to a specific size, but it really has nothing to do with editing latitude...it won't increase it...not in any way that will allow you to recover clipped highlights or shadows. You may have 14.4 stops or DR because the signal (the range of tones between the deepest black and the brightest white) is cleaner...lower error rate per pixel, less noisy.
If you have a 14.4 stop scene that you try to capture with the D800 or D600 in a single frame, you will be clipping 1.2 stops of highlight and/or shadow detail. That 1.2 stops is lost, gone for good, never recoverable...because the SENSOR (i.e. the RAW image) only has 13.2 stops of DR. If you clip 1.2 stops of highlights, pulling down highlights in post will result in clearly blown highlights...you can make them gray rather than white, but those regions will always lack detail. If you clip 1.2 stops of shadows, lifting in post will eventually result in nothing but the stretching of read noise by 1.2 stops, you won't be recovering those lost 1.2 stops of shadows. The D800 is only capable of capturing 13.2 stops of dynamic range in a single shot, as far as editing latitude (the ability to lift shadows without also lifting noise so much that it affects the image) is concerned.
Finally, DXO is making a claim that isn't even backed up by measurements. Print DR is
derived from ScreenDR...it is not measured:
Code:
DRprint = DRscreen * log2 sqrt(originalMpix/referenceMpix)
DXO uses a fixed formula for extrapolating what the noise level should theoretically be in an image downsampled from the original megapixel count of the sensor to the reference megapixel count (3200x2400, or 7.68mp, which is an 8x12" 300ppi 3:2 ratio print). Therefor, a Print DR or 14.4 is not based on actually measuring the dynamic range of a D800 image that was exported at full size from a RAW image, then actually downsampled to 3200x2400 pixels. It's a mathematic extrapolation, not a measure....even though DXO calls it a measure. (I'd be willing to bet
actually measuring the D800's 3200x2400 image would result in a lower dynamic range.)
Personally, this is one of my only real pet peeves with DXO's sensor tests. They
extrapolate,
rather than measure. (The other pet peeve is they
weight...they give out bonuses when a camera "beats" a certain threshold...which skews the differences even more. Two cameras might actually be within the most minimal difference of each other from actual RAW measurements, but because of the fact that bonus weighting comes into play, a camera with a sensor that beats say a DR dB threshold by 0.1dB would appear to be quite a bit better than the competitor, despite the fact that the two are effectively identical as far as the human eye's capacity to observe differences is concerned.)
If you care about editing latitude (which I believe is pretty much THE case here, that everyone is really concerned about editing latitude, given how much the shadow lifting ability of the D800 is brought into the discussion), then you should only concern yourself with DXO's "Screen DR" measure. Screen DR is an ACTUAL MEASURE...it is directly measured from the RAW image data itself, without any extrapolating or resizing or anything else inbetween. It is a REAL value, it is a REAL WORLD value, and it actually tells you something explicit that directly affects an actual post-processing ability you are probably actually concerned about.
Ignore Print DR. It's mathematic magic that doesn't really tell you anything, and is only part of a larger <sarcasm> game of mathematic magic to reduce the complexity of digital cameras to a single number that supposedly tells you everything you need to know about comparing one camera to another on an IQ basis. </sarcasm>
