Messages

1

Canon General / Re: DxOMark vs. Reality

« on: October 12, 2012, 03:29:33 AM »

Yes - but possibly not in the way I think you mean. The original RAW image is spatially oversampled (20MP or more) with respect to the final output resolution (8MP). The process of converting the original high-resolution image to the smaller image is downsampling.

The Wikipedia pages are helpful here, and explain this probably better than I can:

  http://en.wikipedia.org/wiki/Oversampling
  http://en.wikipedia.org/wiki/Downsampling

These also give an example of how this can be exploited to give more resolution (DR) than the actual converter (sensor pixel).

And that must be why an 8KHz downsampling of a CD has more than double the dynamic range of the original, right? You can google theorems, fine, but apparently have no idea what they mean and where they are applicable. So you must work at DXO, then?

Try a test--take a D800 photo of Stouffer wedges, push exposure in post how much you like, and make a note of the darkest one. Now downsample to 1/64, i.e. ~0.5mpix and repeat. How many extra wedges appeared? Exactly as many as how much the *oversampling theorem* is applicable here--i.e. zero.

2

Canon General / Re: DxOMark vs. Reality

« on: October 10, 2012, 10:39:34 AM »

I took this 5 min ago to show the difference between one of  my 5dmk2 and d800

Congratulations, you are the first one to notice that D800 has more DR than 5D2 at base ISO. First prize is a set of tickets for Titanic on it's maiden voyage across the Atlantic.

3

Canon General / Re: DxOMark vs. Reality

« on: October 10, 2012, 07:48:17 AM »

what are DXO´s some would ask.   

Well, they have the word "Science" in their logo and Reichmann doesn't :-)

4

Canon General / Re: DxOMark vs. Reality

« on: October 10, 2012, 07:46:53 AM »

This horse has been beaten to dust. They report 13.2 bits for each pixel. You can gain dynamic range by downsampling. 14.4 stops is based on downsampling to 8mpx.

Are you simple, dishonest or both? The >14-stop-DR fallacy has been discussed multiple times, and has been quite convincingly proven to be just that, a fallacy. You get better SNR from downsampling, but DR cannot magically expand outside the source data boundaries. Get a clue already, please.

5

EOS Bodies / Re: More Big Megapixel Talk [CR1]

« on: October 08, 2012, 03:03:34 AM »

But lets step up a couple of stops. At a signal level of 4, our noise level is 2 (proportional to sqrt of the signal), so at this level we would reduce signal to noise by binning (that is, quantization error isn't the limiting factor).

So you will gain usable dynamic range by increasing resolution, even quantization places a kind of floor on your blackpoint.

Do you understand that on a 14-bit scale, the level 4 corresponds to LV roughly 12 stops down from the maximum? Minimizing noise by pixel binning at LV -12 doesn't do anything to the resultant DR if the per-pixel DR is already over 13 stops. The quantization and 14-bit representation only limit the maximum attainable DR to 14 stops. The clipping of the bottom-end limits it to something <14 stops.

6

EOS Bodies / Re: More Big Megapixel Talk [CR1]

« on: October 08, 2012, 02:53:13 AM »

I don't follow this at all. Luminance isn't negative, so why would it make sense to have negative numbers ? If this clipping really takes place, does this show up on the SNR curves ? I don't really buy that they can inflate the estimated dynamic range by clipping relatively high values (one problem with this is that it leaves some dynamic range on the table).

Negative values in reference to the RAW file blackpoint, which for NEFs is 0. Canon RAWs use either 1024 or 2048. See e.g. here http://theory.uchicago.edu/~ejm/pix/20d/posts/tests/D300_40D_tests/. The values clipped are not "relatively high", they are on the low end of the scale.

Suppose for example our "shadow noise level" (noise at signal level of 1) is 1 -- so 1 on our scale corresponds to the blackpoint.  If we average, theoretically, we could reduce the blackpoint, but our error is stuck at 1 due to quantization. That if I understand it is your argument.

No, it's about the clipping. See the curves in the link I posted. In the case of 40D, the read noise follows normal distribution. You can see that by binning pixels, the resultant noise will converge towards 1024, i.e. the blackpoint. In Nikon's case it would approach something > 0, i.e. a value over the black point set in the NEF.

7

Third Party Manufacturers / Re: 6D Need To Change Something..

« on: October 05, 2012, 12:09:58 PM »

Not sure about the high ISO performance of 6D, but the D600 is outstanding.
Has anyone checked the D600's preview on DPreview, the high ISO performance is better than 5D3 in RAW

Interesting graph. To my eye, the sample frames show an opposite picture (no pun), i.e. D600 shows more noise. The D600 noise is also a lot splotchier for the lack of a better term. Perhaps some sort of on-chip median filter again?

8

EOS Bodies / Re: More Big Megapixel Talk [CR1]

« on: October 04, 2012, 08:05:55 AM »

But I think that is essentially what is going on, because DxOmark's print score is based on a normalization to 8 megapixels. I'd think that would buy you a gain of a stop or so. ( something like log2 ( sqrt(36 megapixels  / 8 megapixels ) )

That's the theoretical part, which would work in an ideal situation with ideal noise characteristics and real numbers instead of quantized integers.

AFAIK, Nikon RAWs currently clamp any digitized negative values to zero (compare that to Canon having a bias value of 2048 in the data). This roughly halves the stdev of the dark frame image captures in Nikon's case, in the end inflating the "measured DR" by roughly one stop.

Another result of this is that for values of low magnitude, oversampling the individual pixel values in SW does not result in the expected behavior of noise converging towards zero. Since the noise-converging-to-zero is a key assumption in the whole "increase-DR-by-binning" scenario, it's quite trivial to notice that the theory doesn't hold water in this case.

9

Lenses / Re: About to buy the 135L, and then saw this....

« on: October 03, 2012, 07:40:12 AM »

As for their lens reviews -- what I don't like about their lens reviews, as I stated previously, is that an attempt to summarize lens performance with a single number (or even three numbers) is likely to be unsuccesful.

That's only a consequence of the real problem. It's their "measurements" that are so incompetent it's almost unbelievable--e.g. the 17-85 kit-lens is "measured" as sharper across the field than the 300L, both mounted on a 7D. Check their resolution maps, MTF charts, whatever, it's all there.

10

Lenses / Re: About to buy the 135L, and then saw this....

« on: October 01, 2012, 03:23:28 PM »

In the case of lenses, they are reporting resolution as a peak measurement - the highest resolution measured at any location in the lens' FoV, at any aperture setting, and for zooms at any point in the focal range.

I think this is only a part of the problem. If you check the measurement data sheet, the 300L maxes out at ~50lp/mm, whereas e.g. the 28-135 reaches or exceeds 70lp/mm across the focal length range. The 70-300 non-L also shows higher lp/mm at 300mm. Utter rubbish, even more so than their sensor shenanigans.

11

Lenses / Re: About to buy the 135L, and then saw this....

« on: October 01, 2012, 01:30:39 PM »

Their lens reviews are rubbish--the 28-135 or 70-300 non-L have more resolution than 300/2.8L II according to them. You'll get more accurate resolution figures by picking out random numbers from a phonebook.

12

EOS Bodies / Re: Who said Canon cameras suck?!?

« on: September 28, 2012, 02:36:01 AM »

One more thing regarding normalization. Suppose your read noise has standard deviation of S. With enough sample data X, the resultant noise will converge to zero. This is the base assumption of phrases such as "doubling resolution increases DR by Y".

However, in the case of D800 and other newer Nikons, there is no bias value in the NEF file data. For the very bottom end of the ADC output, all values for which X+S<0 are truncated to zero. The end result is that normalizing any samples X smaller than S will not any more result in the noise converging to zero, but rather towards (S-X)/2.

This puts a hard limit to the "normalized DR" claims--e.g. no matter how large the set is, the sample value 0 doesn't converge to 0 (which it should in order for the processed output to touch the max DR of 14 stops in this case).

On the other hand, omitting the bias roughly doubles the "measured" DR in the black frame "tests"--half of the read noise is squashed to zero.

13

Animal Kingdom / Re: I'm going squirrely trying to improve these photos (help)?

« on: August 22, 2012, 03:09:53 PM »

I don't usually post pictures, but when I do, they are of squirrels. Here's one solution to busy backgrounds: 500mm+1.4x. I think I went overboard with clarity on this one but it is what it is...

(Info: Cropped to 5x4 + rotated, 7D, f7.1, 1/160, ISO800, hand-held)

14

EOS Bodies - For Stills / Re: DxO results out for 5D3

« on: April 20, 2012, 03:58:57 PM »

If you had a 2 bit monitor would you set 0=bright gray and have the backlight shining through a lot even when fed signal 0? You'd still set 0= liquid crystals block backlight as much as they can = black and then dither to build up lots of intermediate tones and if it had some huge resolution like that you'd build lots of tones.

You can certainly dither between LV+13 and LV+15 in the example system. Not above and not below though. As far as the sensor in this case is concerned, LV+13 *is* black. It is zero in the ADC output scale of [0,3]. Anything below that has ceased to exist after the ADC has digitized it.

15

EOS Bodies - For Stills / Re: DxO results out for 5D3

« on: April 20, 2012, 03:53:14 PM »

Ah, I should have also mentioned that the camera would have 2-bit ADC. Which then means that in this case anything below LV+13 is recorded and interpreted as LV+13. It is a silly hypothetical camera, but the principle is the same for any N-bit ADC--you simply don't have any data beyond the range of the ADC. And no amount of normalization is going to change that.

Clearly, you have a dizzying intellect, but you fail to understand normalization.  Let's define it:

Normalization (ˌnôrmələˈzā sh ən): n. A magical process by which certain data are removed from images and other, nonexistent data are added to images, the net result of which is that those images can be used to support any conclusion in which I firmly believe.

See...it all makes sense, you just had to see it in context.

Ah magick, now I understand!