Will Michael revisit ETTR?

[Guillermo Luijk]

I maintain that ETTR is all about both noise and the number of levels.

In case Emil's example was too abstract, try to decide which of these images has better IQ:



Do you maintain that more levels is better? or having more levels can be irrelevant in presence of noise?.

This is what happens in the lowest stops of a RAW file, lack of levels is irrelevant since noise makes posterization impossible. So having few levels is not an issue, the issue is having noise, and hence doing ETTR is justified only by SNR maximization.

Think of this: if you could have more levels but the same noise thanks to ETTR, the RAW file wouldn't be any better (like in the kid's image), so ETTR wouldn't be of any interest. The only reason for ETTR is reducing visible noise.

That is what happens in ISOless cameras like the K5 and D7000: ETTR through pushing ISO produces more levels in the shadows, but since SNR remains nearly the same on those cameras, ETTR through ISO has no practical advantages in them. More levels, same IQ, highlights possibly clipped.

Regards

[bjanes]

Do you maintain that more levels is better? or having more levels can be irrelevant in presence of noise?.

This is what happens in the lowest stops of a RAW file, lack of levels is irrelevant since noise makes posterization impossible. So having few levels is not an issue, the issue is having noise, and hence doing ETTR is justified only by SNR maximization.

Think of this: if you could have more levels but the same noise thanks to ETTR, the RAW file wouldn't be any better (like in the kid's image), so ETTR wouldn't be of any interest. The only reason for ETTR is reducing visible noise.

That is what happens in ISOless cameras like the K5 and D7000: ETTR through ISO produces more levels in the shadows, but since SNR remains the same on those cameras, ETTR through ISO has no practical advantages in them. More levels, same IQ.

Guillermo,

An excellent demonstration. ETTR through ISO produces more raw levels, but the number of distinguishable levels as defined by Emil remains the same, since it is the exposure that determines noise with these cameras. If one replaces the ETTR dogma with maximizing exposure as Emil suggests, maximal exposure can only be achieved at base ISO. As this thread demonstrates, the number of raw levels is often irrelevant. Michael should restate his ETTR rationale, but by now most readers of this thread should know the true rationale for improved image quality is to maximize exposure.

Regards,

Bill

[Ray]

Emil,
If it is true that the number of levels that all major DSLRs can potentially record, whether in the deep shadows, the low tones or the midtones, is always significantly greater than the number of tones that can actually be distinguished through the noise, then I would have to concede the point that these theoretical numbers of levels at any particular stop in the tonal range are not relevant.

Looking again at your table for the P65+ in reply #85, there seems to be a massive 'overkill' of the number of potential levels in relation to the number of distinguishable tones that are visible through the noise.

top stop = 166 distinguishable tones, 32768 levels
2nd stop= 117 distinguishable tones, 16384 levels
3rd stop= 82 distinguishable tones, 16536 levels
4th stop= 57 distinguishable tones, 8192 levels
5th stop= 39 distinguishable tones, 4096 levels
6th stop= 26 distinguishable tones, 2048 levels
7th stop= 17 distinguishable tones, 1024 levels
8th stop= 10 distinguishable tones, 512 levels
9th stop= 6 distinguishable tones, 256 levels
10th stop= 3 distinguishable tones, 128 levels
11th stop= 2 distinguishable tones, 64 levels
12th stop= 0.9 distinguishable tones, 32 levels

Does this surplus of levels in the P65+ RAW file serve any useful purpose? For example, if the files were 12 bit instead of 16 bit, we would have just one level in the 12th stop and 2 levels in the 11th stop, equal to the number of distinguishable tones in the 16 bit RAW file.

Would it then be likely, if the file were 12 bit, there would be no distinguishable tones in the 12th stop or the 11th stop, or maybe just one tone in the 11th stop? In other words, does a surplus of potential levels, in any given stop, increase to some degree the number of distinguishable tones that may be discerned through the noise?

[Ray]

In case Emil's example was too abstract, try to decide which of these images has better IQ:



Do you maintain that more levels is better? or having more levels can be irrelevant in presence of noise?.

This is what happens in the lowest stops of a RAW file, lack of levels is irrelevant since noise makes posterization impossible. So having few levels is not an issue, the issue is having noise, and hence doing ETTR is justified only by SNR maximization.

Think of this: if you could have more levels but the same noise thanks to ETTR, the RAW file wouldn't be any better (like in the kid's image), so ETTR wouldn't be of any interest. The only reason for ETTR is reducing visible noise.

That is what happens in ISOless cameras like the K5 and D7000: ETTR through pushing ISO produces more levels in the shadows, but since SNR remains nearly the same on those cameras, ETTR through ISO has no practical advantages in them. More levels, same IQ, highlights possibly clipped.

Regards

Guillermo,

Those images are much too small to see any difference, and the current monitor I'm using is not particularly good. At that size, one couldn't tell the difference even between a P&S and a P65+.

By the way, I own a D7000 and have many significantly underexposed images taken at base ISO. Raising exposure always seems to increase the number of distinguishable tones in the shadows.

[Bart_van_der_Wolf]

Does this surplus of levels in the P65+ RAW file serve any useful purpose?

It allows to quantize the noise more accurately, and thus will lead to less posterization risk. Keep in mind that these Raw levels will undergo a gamma adjustment (=boosting shadow contrast, and compressing highlight contrast) before being displayed.

Cheers,
Bart

[Ray]

Here's one of my underexposed D7000 shots, showing the ACR Window before adjustments (except for highlight recovery for some of the bright lights), followed by the processed image.

[ejmartin]

It allows to quantize the noise more accurately, and thus will lead to less posterization risk. Keep in mind that these Raw levels will undergo a gamma adjustment (=boosting shadow contrast, and compressing highlight contrast) before being displayed.

Not sure that this matters much -- stretching or compressing the levels with a gamma adjustment, the noise is stretched/compressed in proportion and its width in proportion to the level spacing remains intact.

There may be some utility to a bit depth of raw that oversamples the data to a degree, as in audio.

[digitaldog]

Guillermo,
Those images are much too small to see any difference, and the current monitor I'm using is not particularly good.

I think that’s the point! Even on the current monitor I’m using which is particularly good.

[Guillermo Luijk]

Those images are much too small to see any difference, and the current monitor I'm using is not particularly good.

Those are synthetic images with the only purpose to didactically demonstrate that more levels not always mean anything useful, they come from the same RAW file. Refer to Emil's numbers to convince yourself about a real case for RAW.

It allows to quantize the noise more accurately, and thus will lead to less posterization risk. Keep in mind that these Raw levels will undergo a gamma adjustment (=boosting shadow contrast, and compressing highlight contrast) before being displayed.

What is better, a computer than can perform a task in one milisecond, or a computer that can do it in one microsecond?. For a human being, both are as good, no matter if the second computer was 1000 times faster.

If posterization is not an issue with any arbitrary exposure (and this is the point of the whole story), it's irrelevant that any higher exposure could theoretically provide less posterization risk.

BTW, if your scene contains a plain colour area (like a sky), the more you expose it, the less noise it will have, and hence the more posterization risk when converting to 8-bit JPEG. I have experimented this issue on HDR interiors, where walls absolutely clean of noise, which sounds like a good idea at first, because of RAW overexposure are quite prone to display bands when converted to JPEG. So funily ETTR could lead more easily to posterization in some cases than a lower exposure (posterization in the highlights due to the absence of noise, never due to the lack of levels in the RAW capture).

I already posted an example of this; no one has posted so far a practical example of the opposite, the Myth of ETTR levels!!!.

Regards

[bjanes]

Here's one of my underexposed D7000 shots, showing the ACR Window before adjustments (except for highlight recovery for some of the bright lights), followed by the processed image.

And what does this mean? It shows that ETTR is not that necessary with a camera having good noise characteristics. With this camera, slight underexposure may be better than attempted ETTR with clipping of highlights.

Regards,

Bill

[Ray]

And what does this mean? It shows that ETTR is not that necessary with a camera having good noise characteristics. With this camera, slight underexposure may be better than attempted ETTR with clipping of highlights.

Regards,

Bill

Bill, It's a bit puzzling that DXO Mark attribute almost a 14 stop DR for the D7000, and a full 14 stops DR for the K5, yet in a 14 bit RAW file there would only be one level available for this 14th stop, 2 levles for the 13th stop, and 4 levels for the 12th stop.

The image of a hotel lobby in Thailand posted above, although apparently considerably underexposed, could be considered an ETTR in relation to the specral highlights of the lamps. Therefore, it seems reasonable that a significant part of that particular image will fall within the 14th to 12th stops.

Are 1 to 4 levels sufficient for the D7000 to make the most of these deep shadows, despite the noise?

[kwalsh]

Bill, It's a bit puzzling that DXO Mark attribute almost a 14 stop DR for the D7000, and a full 14 stops DR for the K5, yet in a 14 bit RAW file there would only be one level available for this 14th stop, 2 levles for the 13th stop, and 4 levels for the 12th stop.

It isn't unusual for DxO to be puzzling since their measurements while useful for comparing sensors to a degree use absolute metrics that really aren't photographically relevant.

For them DR means 0dB SNR in an image resized to 8MP.  For the K5/D7000 with a 16MP sensor that means the SNR will be 3dB lower once resized (or equivalently that the DR will be 0.5 EV better once resized, you can see this effect on DxO if you click the "Screen" tab instead of the default "Print" tab, the "Screen numbers will be 0.5EV lower).  The quantization noise floor, assuming the ADC is almost saturated when the pixel well is full, would be 6.02dB*(14 bits) + 4.77dB = 89dB at the native 16MP resolution.  At 8MP resized it would be 92 dB.  Converting back to EV we get 92dB/6.02dB = 15.3 EV as the theoretical maximum DR from a perfect 14-bit ADC perfectly loaded on a 16MP imager using DxO's methods.

The problem is that no photographer is going to find 0dB to be acceptable shadows at all.  The practical DR is lower than DxO's 14 EV.  They had to chose an SNR level for reference so they used 0dB which is a nice engineering number and what sensors designers usually use to specify pixel DR.

That's a lot of math gobbledygook that probably really isn't worth following as far as helping anyone take nice pictures.  The take away is that in general it is best to use the DxO measurements as a relative comparison tool between cameras.  The absolute numbers and reference levels they use are a bit arbitrary as far as practical photography goes.

Ken

[bjanes]

It isn't unusual for DxO to be puzzling since their measurements while useful for comparing sensors to a degree use absolute metrics that really aren't photographically relevant.

For them DR means 0dB SNR in an image resized to 8MP.  For the K5/D7000 with a 16MP sensor that means the SNR will be 3dB lower once resized (or equivalently that the DR will be 0.5 EV better once resized, you can see this effect on DxO if you click the "Screen" tab instead of the default "Print" tab, the "Screen numbers will be 0.5EV lower).  The quantization noise floor, assuming the ADC is almost saturated when the pixel well is full, would be 6.02dB*(14 bits) + 4.77dB = 89dB at the native 16MP resolution.  At 8MP resized it would be 92 dB.  Converting back to EV we get 92dB/6.02dB = 15.3 EV as the theoretical maximum DR from a perfect 14-bit ADC perfectly loaded on a 16MP imager using DxO's methods.

The problem is that no photographer is going to find 0dB to be acceptable shadows at all.  The practical DR is lower than DxO's 14 EV.  They had to chose an SNR level for reference so they used 0dB which is a nice engineering number and what sensors designers usually use to specify pixel DR.

That's a lot of math gobbledygook that probably really isn't worth following as far as helping anyone take nice pictures.  The take away is that in general it is best to use the DxO measurements as a relative comparison tool between cameras.  The absolute numbers and reference levels they use are a bit arbitrary as far as practical photography goes.

Ken

Ken,

It is quite true that a SNR of 0 db is not useful for photography.  As you suggest, this is the engineering dynamic range. The photographic range will be less, but that range depends on how much noise you can tolerate in the shadows. The DXO full SNR plot can be used to determine the photographic DR for any given noise floor, as explained by Emil Martinec in this post.

Regards,

Bill

[Ray]

Those are synthetic images with the only purpose to didactically demonstrate that more levels not always mean anything useful, they come from the same RAW file. Refer to Emil's numbers to convince yourself about a real case for RAW.

I see! So you've deliberately created a set of artificial circumstances to demonstrate a point which may be valid only within that set of artificial circumstances. Is that correct?

[Ray]

The practical DR is lower than DxO's 14 EV.  They had to chose an SNR level for reference so they used 0dB which is a nice engineering number and what sensors designers usually use to specify pixel DR.

That's a lot of math gobbledygook that probably really isn't worth following as far as helping anyone take nice pictures.  The take away is that in general it is best to use the DxO measurements as a relative comparison tool between cameras.  The absolute numbers and reference levels they use are a bit arbitrary as far as practical photography goes.

Ken

Really! The two images below, taken with the D7000 shortly after I received it, indicate that the 14th stop contains useful, althought very limited, visible detail. At least you can see that what I've photographed is a Dynamic Range Chart, and make out a few of the largest numbers.

Perhaps my maths is wrong, but I calculate that a range of exposures from 4 seconds to 1/2000th represents 14 stops.

Despite the fact that the target has been artificially created (by Jonathan Wienke), the shots are still 'real-world' shots.

The shot in the 14th stop seems to me to contain too much detail for a 1 level capacity. So maybe my assumed ETTR exposure at 4 seconds is in fact an overexposure.

[ejmartin]

The shot in the 14th stop seems to me to contain too much detail for a 1 level capacity. So maybe my assumed ETTR exposure at 4 seconds is in fact an overexposure.

One level per pixel is many levels when averaged over many pixels.  S/N is scale dependent, this is why DxO's figures differ for 'screen' and 'print' for example.  So one level is sufficient for a single pixel, it will yield more than one distinguishable level in a downsampled image such as the one you are displaying.  (For example group four pixels; the average of a level which can be zero or one for each pixel now ranges over 0,1/4,1/2,3/4,1 so there are five levels, however S/N only doubles and so only two are distinguishable from noise on average; the analysis scales in a straightforward way to coarser scales.)

[kwalsh]

Really! The two images below, taken with the D7000 shortly after I received it, indicate that the 14th stop contains useful, althought very limited, visible detail. At least you can see that what I've photographed is a Dynamic Range Chart, and make out a few of the largest numbers.

And most photographers would choose a black level to make all that noise disappear and lose the detail.  That's all I meant by practical, not that you couldn't see anything.  In fact, with a large enough structures you should be able to see features well below 0 dB.

Perhaps my maths is wrong, but I calculate that a range of exposures from 4 seconds to 1/2000th represents 14 stops.

Seems right to me.

The shot in the 14th stop seems to me to contain too much detail for a 1 level capacity. So maybe my assumed ETTR exposure at 4 seconds is in fact an overexposure.

I wouldn't be surprised if your exposures are correct.  There should still be details this large visible in an image with a bit depth of one.  After all, this very small image uses only 1 bit per pixel and I'm sure we can all tell exactly what it is:



Also worth remembering most every digital audio output device these days uses a 1 bit digital to analog converter.  One bit per sample is plenty of information in the right use and context.

Ken

[Ray]

And most photographers would choose a black level to make all that noise disappear and lose the detail.  That's all I meant by practical, not that you couldn't see anything.  In fact, with a large enough structures you should be able to see features well below 0 dB.
Ken

I understand that subjective preferences may determine the choice of black level, but I don't understand how one can see features below 0dB.

Seems right to me.

I wouldn't be surprised if your exposures are correct.  There should still be details this large visible in an image with a bit depth of one.  After all, this very small image uses only 1 bit per pixel and I'm sure we can all tell exactly what it is:

That's a good point. But my concern here is about the detail provided by a given number of levels. Would that parrot appear more detailed if we raised the number of levels to 4 bit, despite any noise?

[Ray]

One level per pixel is many levels when averaged over many pixels.

Makes no sense to me. One is one, and the avarage of one, is one. In a context of one level, any pixel is either black or white. However, in the deep shadows there can be no white, so one level is either black or near black.

Is this not correct?

I'm deducing that the black equlas 0dB and the near-black is the one level.

[kwalsh]

I understand that subjective preferences may determine the choice of black level, but I don't understand how one can see features below 0dB.

You can see features below 0dB if they are large enough.  It is similar to SNR improving when you downsize an image.  Your visual system is actually surprisingly good at picking out large patterns masked by fine patterned noise.  In this case I meant 0 dB in the DxO context - pixel SNR of 0 dB in an 8MP image.  You won't see any fine pixel level detail at say -10 dB but if we have big feature, say one taking up nearly a quarter of the image, that will be visible.

This is the whole issue with terms like "0 dB" or "1-bit", they don't really mean anything on their own unless we have some other context (total number of pixels, size of detail of interest, etc. etc.).

That's a good point. But my concern here is about the detail provided by a given number of levels. Would that parrot appear more detailed if we raised the number of levels to 4 bit, despite any noise?

You'd have to leave out the "despite any noise" part to get a useful answer.  Noise is the key here.  If there was a 4-bit noise free image certainly finer detail would be visible in the 4-bit image than the 1-bit image.  If the image was already very noisy the extra bits wouldn't do any good.  That is the case with the K5 though - there is already read noise there larger than the quantization noise of the ADC so extra bits won't buy you anything. 

Finally, just to add to the confusion, converting from 4-bits to 1-bit we'd add noise before converting to 1-bit to get the best detail in the final 1-bit image! (look up dithering for the details)

Ken