> 100 point DXOMark Sensor score.

[MrSmith]

That's your interpretation not mine.
As I said above i was commenting on the nature of web chatter about stuff they know little about and are never likely to have in their hands,
 'it scores 100 on Dxo so it's the best camera in the world and your camera sucks'
The guys in lab coats are just supplying the measurebators ammunition (and more importantly some data for those who know how to look beyond a perfect 100 score)

[RobertJ]

Hopefully we can retire the term "measurebator."  It's anti-intellectual.

Since when?  I seem to recall many people on this forum and around the web using the term.  It's no more ridiculous than the term "tweet."  Look it up and get with the program.  Welcome to the internet son.

Also what MrSmith says is correct.  His comment is the most valid on this entire thread.

RED makes some neat products, but the Arri Alexa will remain the gold standard.  Look at ALL NINE of the Academy Award Nominated films for Best Picture.  NONE OF THEM... let me repeat that.... NONE OF THEM used a RED camera.  Many of them used Arri film or digital cameras. 

No Dragon, no Epic, no Scarlet, no nothing.

[LKaven]

Since when?  I seem to recall many people on this forum and around the web using the term.  It's no more ridiculous than the term "tweet."  Look it up and get with the program.  Welcome to the internet son.

The word "measurebator" is made up of which two English words?  The implication being?  I thought so.  The word "tweet" compares ... how?

We're going to kill this meme off, but good.

I remember when people -- everyone -- used the N word -- every day all the time -- to refer to blacks.  I looked that one up and refused to get with the "program."

And I'm not your "son", unless you and "mom" want to help me out with a new car.

[UlfKrentz]

That much is plausible ... once one acknowledges that the DR has probably been enhanced by the averaging of information from several frames taken in rapid succession (temporal noise reduction), sort of like HDR on the fly. snip

I don´t think this is based on frame averaging. DSMC cameras have a HDR mode to do this and this way reach for at lot more DR than shown here.
That said, a RED won´t become my daily stills camera for a couple of reasons but we already used it in parallel with a MDB for a portrait and decided to go with the R3Ds for the final output, go figure. There has been a comparison among cameras of member CB, and that was only the "MX". I don´t like the Red fanboy hype at all and I would never think of just pulling a frame from motion work but if you set up your workflow for a still you might use a motion camera and end up with a good result (and awful motion).

I couldn´t care less who won the best picture awards for what reasons, RED is a very young company and they stirred up some dust in this traditional business.

Cheers,

Ulf

[hjulenissen]

Welcome to the internet son.

Being rude on the internet may be common. This does not mean that it should be applauded.

-h

[BJL]

I don´t think this is based on frame averaging. DSMC cameras have a HDR mode to do this and this way reach for at lot more DR than shown here.

Temporal averaging of consecutive frames is not the same as the usual HDR; it takes several consecutive frames all at the same exposure level and averages them, rather than the usual HDR approach of using frames at very different exposure levels.  So it will not produce the big DR expansion of HDR, but it will reduce the noise floor, particularly when using in a stationary test subject so that all frames see exactly the same light at each photosite.  Roughly, if several frames are averaged with equal weight (not that this is exactly what RED does), the level of random uncorrelated noise will go down and the SNR and DR go up by the square root of the number of frames averaged, so if two consecutive frames are averaged, the improvement is half a stop, four frames improves by one stop, and so on.  Color depth will improve in the same way.

[UlfKrentz]

Ok, I understand the technique (behaves like having more transistors in parallel of an input stage of an amp) what I don´t understand is where they assume multiple frames come from. Actually RED announced an ADD tech for pulling stills of a higher quality with their new release of RedCineX and Dragon files, which may take advantage of this method, they also have a new OLPF now, which both were not involved in this test as far as I can see. Are you talking about multiple readouts during the selected "shutter" speed? Anyway, used to MFDB the DR of a Red camera (even MX) is impressive. Obviously the rest is comparing apples and oranges ;-)

Cheers,

Ulf

[BJL]

Ok, I understand the technique (behaves like having more transistors in parallel of an input stage of an amp) what I don´t understand is where they assume multiple frames come from.

I see good evidence for temporal averaging in the DXO data (as do the folks as DXO):

1.  The total SNR graphs go up to about 50dB (at minimum ISO exposure index, 100% illumination).  With just photon shot noise, that requires counting at least 100,000 photons (the photon count needed to achieve a given SNR in dB, SNRdB, is 10^(SNRdB/10).)
That is about 4000 photon counts per square micron of photosite area, more than twice what is typical and almost three times what I see measured for the D800 which has similar photosite size.
That can only be achieved by some combination of (a) deeper wells, to actually count 4000 photo-electrons per square micron, and (b) combining the photo-electron counts from several frames (temporal averaging).

2.  The base ISO speed Ssat (what DXO calls "ISO") is 104, comparable to and indeed a bit higher than cameras like the D800 [Ssat=75 for the D800].

3. If there were no temporal averaging, having over double the well depth of the D800 and also this higher Ssat would require almost tripling the quantum efficiency. (Doubling well depth at equal QE would half the base ISO speed Ssat.)

4. Sensors like that of the D800 are close to the maximum possible QE for a Bayer CFA sensor, and triple the QE of the D800 would be beyond 100%: it seems impossible that RED could have increased the QE of its Bayer CFA sensors by nearly enough to explain the DXO measurements.


By process of elimination, temporal averaging seems almost certain.


Arguably, this still gives a legitimate advantage for exposures when light is abundant: effectively, the light used to form each frame is being gathered over a longer time than sensor saturation normally allows, so it is almost equivalent to increasing the well depth and decreasing the base ISO speed Ssat by a factor of more than two -- and yet doing this with only about 1/24s between frames using RED's fast rolling shutter, and so handling subject and camera motion better than frame averaging with a normal still camera, which would need to have more time between the frames.

So a stills photographer might get some advantage when longish total exposure times are acceptable, in exchange for paying about twenty times as much as for a stills cameras with the same sensor size.

[Bart_van_der_Wolf]

By process of elimination, temporal averaging seems almost certain.

Indeed, and it satisfies Occam's Razor.

Cheers,
Bart

[LKaven]

I suspect that they are doing temporal averaging, but not of equal length exposures. 

I suspect they are taking the second shot at a much faster shutter speed.  In doing so, they can fit two exposures into a what the users expects would be the space for one (e.g., 1/60th plus 1/250 or so all in the space of 1/50th).  And the clear win on dynamic range is at the top end, with some improvement in the midtone response. 

Highlight protection is what a cinema shooter wants here I think.

[BJL]

I suspect that they are doing temporal averaging, but not of equal length exposures. 

I suspect they are taking the second shot at a much faster shutter speed. 

Since it is from a video feed, I guess a weighted rolling average of "equal" frames: each frame gets the same exposure, and the processed output at each time stamp is something like:
1/4 the previous frame + 1/2 the current frame + 1/4 the next frame.

But maybe someone who actually knows something about how temporal averaging is typically used in video processing could correct my guesswork?

[LKaven]

Since it is from a video feed, I guess a weighted rolling average of "equal" frames: each frame gets the same exposure, and the processed output at each time stamp is something like:
1/4 the previous frame + 1/2 the current frame + 1/4 the next frame.

But maybe someone who actually knows something about how temporal averaging is typically used in video processing could correct my guesswork?

Hmm, hadn't thought of that.  That's interesting. 

But it seems there'd be a lot of detail lost on moving subjects.  I wonder if this is good enough for cine?  If you do one long and one very short exposure, you can use the very short exposure as a basis for detail enhancement.  And you get the extra stops of headroom at the top end, and clean.  Enhancing midtone response is a nice dividend. 

[UlfKrentz]

BJL,

Thank you for your in-depth explanation. I have to admit I also like the artistic side of our job.
 

Cheers,

Ulf

[Bart_van_der_Wolf]

Hmm, hadn't thought of that.  That's interesting. 

But it seems there'd be a lot of detail lost on moving subjects.  I wonder if this is good enough for cine?  If you do one long and one very short exposure, you can use the very short exposure as a basis for detail enhancement.  And you get the extra stops of headroom at the top end, and clean.  Enhancing midtone response is a nice dividend. 

Hi Luke,

The problem though, besides complexity, is that equal exposures lead to lower average shot noise. It's also easier to average (1 bit shift per 2 exposures). Also, it's easy enough to store the cumulative result of multiple exposures and discard read noise at the same time, even within 16-bits (4x 14-bits <= 16-bits).

Cheers,
Bart

[LKaven]

Hi Luke,

The problem though, besides complexity, is that equal exposures lead to lower average shot noise. It's also easier to average (1 bit shift per 2 exposures). Also, it's easy enough to store the cumulative result of multiple exposures and discard read noise at the same time, even within 16-bits (4x 14-bits <= 16-bits).

There's a good line of reasoning in it, and I don't discount it.  But I was also thinking that with two exposures, one two stops faster in shutter speed, you'd get both the extra highlight headroom, and almost all of the benefits of averaging (after normalizing) everywhere but in the lowest 2-3 stops.  

I was thinking that the benefits of highlight headroom and added midtone response would be most desirable to a cinematographer, and that the shadow response would be secondary.  After all, there are some natural noise reduction effects in rapid presentation of sequential images.  People do not perceive shot noise going by at 30 frames per second so readily.  

Then I was thinking of the benefits of detail enhancement, which could be varied.  I was comparing this to the inevitable smearing that would happen borrowing from adjacent frames, which seems like a compromise best suited for broadcast and streaming applications.

Having said all that, you might be right.  Temporal noise reduction could be strictly an interframe function, and that might be all that's going on.  But that would be a little disappointing.  And a little disappointing of DxO to classify this in the same category as single-shot cameras.  If Nikon were to introduce, in firmware, a two-shot burst mode with in-camera averaging, could they too take the lead back in the sensor rankings?

[UlfKrentz]

snip  If you do one long and one very short exposure, you can use the very short exposure as a basis for detail enhancement.  And you get the extra stops of headroom at the top end, and clean.  Enhancing midtone response is a nice dividend. 
[/quote]

Luke,

That´s exactly what RED does in HDR mode. You just select "HDR" and dial in the number of stops of extra highlight protection. The camera than records two video tracks simultaneously which can be blended in RedCineX in various ways. Downside is the doubled amount of data, the need of higher compression and with fast moving objects different blur and timing caused by two exposures and different shutter speeds. Great onboard feature though.

Cheers,

Ulf

[BJL]

... If you do one long and one very short exposure, you can use the very short exposure as a basis for detail enhancement.

That´s exactly what RED does in HDR mode. You just select "HDR" and dial in the number of stops of extra highlight protection.

Indeed, that is how RED's HRDx mode works, but DXO did not that special mode for its test, so all frames in the video stream fem which the stills were derived are at equal exposure. In this normal mode, the 24fps output is produced from 24fps capture, so no "two frames captured per each frame output". That is why a rolling average is the only way I can see it operating in the mode use for the DXO testing.

To answer a question that someone (Bernard?) asked; I would think that a digital still camera could add a similar "DR enhancement" mode, but it would require either
1) the camera held dead steady, on a tripod
or
2) a very short lag between the frames, which would probably require a video mode that outputs full resolution, with no sub-sampling or binning done on the sensor chip.
Video-stills hybrid models like the Panasonic GH4 might have the latter hardware capability.

[Abstraction]

I have been a publishing poet and writer and a serious photographer for forty years.

All the information one can get abut ones art is good.

An artist is a person on whom no information is lost.

[barryfitzgerald]

Few will care about numbers. It might help if DxO bothered to tell you how they test and what exactly they are testing.
Based on my own personal use of quite a few cameras I consider DxO to be largely irrelevant and in some cases I've had serious doubts about their methods and scores.

The lens reviews are just as vague (perceived mp ratings for lenses)

There may be some value in some of their results, but they've gone about it the wrong way and are unconvincing and pretty much pointless to "people who actually take photos rather than measure things"