Will Michael revisit ETTR?

[joofa]

Guillermo, I must say that I am very pleased at your measured, sensible, and polite responses. Thanks for all the discussion.

From both, real images of flat patches.

You mean an image acquired of a flat patch?

No idea how the formula will change for other type of rescaling, I just chose this criteria because the only intention was to compare sensors. With other type of rescaling the change in SNR will be different of course, but for typical interpolation methods I'm pretty sure this model works well.

Okay that is fine, as long as you have an appreciation of the formula changing due to a change in the resampling method.

Sincerely,

Joofa

[Guillermo Luijk]

So there are two reasons:
1. To avoid noise affecting the (effectively) estimation of signal in isolation
2. To better emulate the human perception

Correct!.

whether Michael's explanation was 100% accurate, etc... is essentially irrelevant because there is only one exposure that is ideal for the sensor, the one that puts the maximum of photons in the best zone.

I dont' fully agree. Reading Michael's article, one might think there are two practical advantages derived from ETTR:
1. Less visible noise
2. More levels

So even if noise is not a problem in your application at a conventional exposure, one could insist in ETTR thinking he will get some other advantage. But this is not true since those extra levels don't mean any practical benefit, so the conclusion is that if noise is not an issue in your application, insisting in ETTR is a complete waste of time.

[Guillermo Luijk]

You mean an image acquired of a flat patch?

Yes. A patch only means obtaining a pair (exposure, SNR). Several patches allow to plot the SNR curve.



Signal in the patch is the average value on it and hence the patch exposure. Noise in the patch is the patch's stdDev. Having S and N, we have the desired (S, SNR).

The blurring of real images was just to obtain the S value (local average exposure). The SNR value is then obtained from the curve for the X-axis value indicated by the previous S.

[joofa]

Yes. A patch only means obtaining a pair (exposure, SNR). Several patches allow to plot the SNR curve.

Are we back to square one? A real image (of your cat, dog, cat, people, cat, landscape, cat, ..., cat) is not always easy to segment into patches as you have shown in the image in your example above. Do you agree with this statement?

Sincerely,

Joofa

[hjulenissen]

Correct!.

At very large signal levels, the SNR should be large. Then noise should not affect the estimation of signal (and thereby SNR). The likelihood that a "code 20" signal is changed into a code 16211 by noise should hopefully be very, very low.

One problem with including perception: if we (primarily) sense the world limited to e.g. 1 MP, then any camera/raw converter that does crude NR by downsampling would appear to have better SNR. Is this sensible?

So even if noise is not a problem in your application at a conventional exposure, one could insist in ETTR thinking he will get some other advantage. But this is not true: if noise is not an issue in your application, insisting in ETTR is a complete waste of time.

Even if noise is not considered a problem at the time of capture, I think that some would prefer to capture it as accurately as possible (who knows what post-processing will be applied many years into the future).

Is this based on self-dithering properties of noise, or some other post of yours that I did not comprehend?

-h

[Guillermo Luijk]

Are we back to square one? A real image (of your cat, dog, cat, people, cat, landscape, cat, ..., cat) is not always easy to segment into patches as you have shown in the image in your example above. Do you agree with this statement?

Of course, but we don't need to segment it in patches, just would calculate each pixel's surrounding exposure (that's the blur filter).

Another option is to segment the image in wide EV zones, and consider a SNR for each whole EV area. But again we need some kind of averaging; otherwise noise and textures could fool us:



This image was noiseless, since it was produced with a blending of several exposures, but still textures in the ground make it difficult to clearly differentiate the EV borders.

[bjanes]

Are we back to square one? A real image (of your cat, dog, cat, people, cat, landscape, cat, ..., cat) is not always easy to segment into patches as you have shown in the image in your example above. Do you agree with this statement?

Joofa,

Your past few posts have been argumentative and unhelpful. One can not give a SNR for an entire image. Each area will have a different SNR according to the luminance of the area being studied. The highlights will have a high SNR and the shadows a low SNR. Look at the SNR plot by DXO for the D7000 for base ISO. If the image is fully exposed to the right, the highlights will have a SNR of about 45 dB. The deep shadows will have a lower SNR.

Regards,

Bill

[Guillermo Luijk]

Even if noise is not considered a problem at the time of capture, I think that some would prefer to capture it as accurately as possible (who knows what post-processing will be applied many years into the future).

But you must think ETTR is not for free. You need longer exposure times (that may cause a not so sharp image), or wider apertures (losing DOF) or higher ISO (clipping highlights information). In general, once noise requirements are met, I prefer to use the fastest shutter I can afford to optimise sharpness.

[joofa]

Joofa,

Your past few posts have been argumentative and unhelpful.

Sorry Bill if you thought so.

One can not give a SNR for an entire image. Each area will have a different SNR according to the luminance of the area being studied. The highlights will have a high SNR and the shadows a low SNR. Look at the SNR plot by DXO for the D7000 for base ISO. If the image is fully exposed to the right, the highlights will have a SNR of about 45 dB. The deep shadows will have a lower SNR.

I think I already answered this here:

http://www.luminous-landscape.com/forum/index.php?topic=56947.msg464138#msg464138

Sincerely,

Joofa

[hjulenissen]

But you must think ETTR is not for free. You need longer exposure times (that may cause a not so sharp image), or wider apertures (losing DOF) or higher ISO (clipping highlights information). In general, once noise requirements are met, I prefer to use the fastest shutter I can afford to optimise sharpness.

So ETTR it is a sort-of single-sided requirement: if circumstances allow, increase exposure until nearly clipping. If not, do whatever compromise between motion-blur, DOF and noise that satisfy your artistic requirements.

There are cases where the problem is small aperture due to bright sunlight and unwanted diffraction, perhaps using flash for evening out the light (and limiting the minimum exposure time), and the lack of ND filtering or lower sensitivity ISO modes. In that case, knowing exactly how "hot" you can expose the sensor using ETTR-inspired techniques and knowing that it will only gain the SNR seems like a good thing to know.

-h

[PierreVandevenne]

So even if noise is not a problem in your application at a conventional exposure, one could insist in ETTR thinking he will get some other advantage. But this is not true since those extra levels don't mean any practical benefit, so the conclusion is that if noise is not an issue in your application, insisting in ETTR is a complete waste of time.

On the whole I agree - but saying "if noise is not an issue" essentially means that the DR of the scene as you want to capture it fits nicely in the DR window of your camera. And if it fits, but at the bottom, fitting it at the top will in most cases give a bit more latitude for post processing. Also agree on the cost of ETTR, although I found out with practice that using ETTR-Light (overexposing slightly and systematically) with my Canon cameras is 99.9% of the time effective. It reverses the effect of the defensive defaults chosen by the Canon engineers.

 

[joofa]

Another option is to segment the image in wide EV zones, and consider a SNR for each whole EV area. But again we need some kind of averaging; otherwise noise and textures could fool us:

That is one way to proceed and fine with me. A point of the whole exercise was how to figure out that "some kind of averaging", and how to separate noise from texture, as you have also identified. That is not always an easy task to do.

Sincerely,

Joofa

[PierreVandevenne]

Joofa,

Your past few posts have been argumentative and unhelpful.

I wouldn't say so. He raises some valid points. While Emil's work is extremely impressive and beyond reproach, some of its results, especially in the normalization of SNR vs Resolution area, should not be extended to extremes by means of black boxes (resize operations in software for example, which are not equivalent to hardware binning). But it is again a long topic of little practical interest for photography.

[Schewe]

But you must think ETTR is not for free. You need longer exposure times (that may cause a not so sharp image), or wider apertures (losing DOF) or higher ISO (clipping highlights information). In general, once noise requirements are met, I prefer to use the fastest shutter I can afford to optimise sharpness.

Just to be clear, if one is already shooting on a tripod (useful to get maximum image sharpness) the penalty for slowing down the shutter speed isn't great. Still a factor for subject motion blur, yes...but not for camera shake. If you are shooting hand held, yes, the focal length and shutter speed will be a factor that must be weighed and decided about...it's far easier to use noise reduction to mitigate shot noise than try to use a sharpening effect to correct for camera shake.

There are a lot of factors that must be evaluated when making a capture...but the science still shows that more photons is always better when it comes to noise. When to use ETTR and when it'll help the image quality is for the photographer to decide, scene by scene and image by image. Above all else, take the shot...don't let all the various factors that must be considered keep you from pressing the shutter. Take the shot...

[Bryan Conner]

There are a lot of factors that must be evaluated when making a capture...but the science still shows that more photons is always better when it comes to noise. When to use ETTR and when it'll help the image quality is for the photographer to decide, scene by scene and image by image. Above all else, take the shot...don't let all the various factors that must be considered keep you from pressing the shutter. Take the shot...

Right on!  TTFP....Take The F***ing Picture!

[bjanes]

I wouldn't say so. He raises some valid points. While Emil's work is extremely impressive and beyond reproach, some of its results, especially in the normalization of SNR vs Resolution area, should not be extended to extremes by means of black boxes (resize operations in software for example, which are not equivalent to hardware binning). But it is again a long topic of little practical interest for photography.

As Emil correctly points out, software averaging of 4 pixels gives twice the SNR of using only one pixel, since noise adds in quadrature. Hardware binning is frequently used in scientific applications with monochrome sensors, but is much more difficult to do with Bayer array sensors. The only color cameras for general photographic use AFAIK that use hardware binning are the Phase One Sensor Plus series. Hardware binning of 4 pixels gives 4 times the SNR, since only one read noise is involved, whereas 4 read noises occur with software pixel averaging. See the explanation by an engineer on the Phase One site (click on the IQ Sensor+ tab).

Noise averaging is of great importance in photography. If one holds sensor size constant and doubles the pixel count, the pixels are smaller and have a lower SNR per pixel, but the overall SNR will not be affected that much. When comparing cameras with different pixel counts for a given print size, some type of normalization must be performed as in the DXO screen and print data.

Regards,

Bill

[joofa]

As Emil correctly points out, software averaging of 4 pixels gives twice the SNR of using only one pixel, since noise adds in quadrature.

Unfortunately, this statement is technically incorrect in general in the context of images, as pointed out several times.What you said will only work if the signal is constant (flat patch). Think about it like this: even if we don't worry about how noise is being added, when you add pixels you are changing (usually blurring) the pixels also, so the signal has changed. Recall, SNR is signal to noise ratio, so SNR changes differently than "4 pixels gives twice the SNR of using only one pixel". I would invite Emil to do calculations himself to verify this fact rather than theoretical arguments.

Sincerely,

Joofa

[bjanes]

Unfortunately, this statement is technically incorrect in general in the context of images, as pointed out several times.What you said will only work if the signal is constant (flat patch). Think about it like this: even if we don't worry about how noise is being added, when you add pixels you are changing (usually blurring) the pixels also, so the signal has changed. Recall, SNR is signal to noise ratio, so SNR changes differently than "4 pixels gives twice the SNR of using only one pixel". I would invite Emil to do calculations himself to verify this fact rather than theoretical arguments.

I will await Emil's response, since he is more technically adept than myself. In the meantime, what are your calculations?

Regards,

Bill

[bjanes]

Unfortunately, this statement is technically incorrect in general in the context of images, as pointed out several times.What you said will only work if the signal is constant (flat patch). Think about it like this: even if we don't worry about how noise is being added, when you add pixels you are changing (usually blurring) the pixels also, so the signal has changed. Recall, SNR is signal to noise ratio, so SNR changes differently than "4 pixels gives twice the SNR of using only one pixel". I would invite Emil to do calculations himself to verify this fact rather than theoretical arguments.

While awaiting Emil's analysis (presuming he takes the trouble to reply to your post), the DXO normalization procedure is of interest. The normalized SNR equation is:

    SNRnorm = SNR + 20 * log10 (sqrt[N/N0]),

where N0 is the original number of pixels, N is the number of pixels for the sensor with the higher pixel count, and SNR is the original SNR.

If we average 4 pixels into one, the formula shows that the SNR increases by 6.02 dB or 1 stop, in agreement with Emil's figure. I think DXO is using flat patches to derive their figures. As you suggest, in real world use with demosaiced images, the SNR may be somewhat less than the theoretical value. The DXO engineer also states that 4:1 binning outside the sensor hardware doubles the SNR whereas hardware binning quadruples the SNR. What are your figures?

Regards,

Bill

[Guillermo Luijk]

the DXO normalization procedure is of interest. The normalized SNR equation is:

    SNRnorm = SNR + 20 * log10 (sqrt[N/N0]),

where N0 is the original number of pixels, N is the number of pixels for the sensor with the higher pixel count, and SNR is the original SNR.

If we average 4 pixels into one, the formula shows that the SNR increases by 6.02 dB or 1 stop, in agreement with Emil's figure. I think DXO is using flat patches to derive their figures. As you suggest, in real world use with demosaiced images, the SNR may be somewhat less than the theoretical value.

I think the 'patches discusion' is going farer than it really deserves. Averaging 4 pixels of the same value + their individually added noise improves SNR by 2, I think we all agree here. If the signal is different on each pixel, the noise will also be different, so SNR will be different on each pixel. If each source pixel has its own SNR, it's nonsense to look for 'SNR improves by a factor of X', since source SNR is not unique.

But once defined, the 'patch model' can be extended to any real world case with the purpose of SNR normalisation. It will always be an approximation, but a valid approximation to make noise performance from different sensors comparable. It is not intended to quantify SNR on any real picture of our cat.

That DxOMark's formula was already posted in the thread as derived from quadrature noise addition:

SNR_norm_dB = SNR_perpixel_dB + 20 * log10 [(Mpx / 12,7)^0.5]

Regards