Exposing to the Right - Revisited

[Graystar]

I needed some time to digest all this, but I'm not so sure that we can ignore the 'RAW levels' that easily. I agree that there's a lot going on in the demosaicing step (although not all algorithms are very complicated), but the problem of clipping doesn't usually occur on isolated pixels. In most cases there's a region of several adjacent pixels that gets overloaded, and that can not be corrected by interpolation within a colour channel.

But that's not related to the "levels" issue. The levels issue is related to the shadow end, where fewer and fewer bits are used to describe the luminance level. So on a 12-bit camera, when you get 10-stops down from the max signal level you only have 3 values to describe an entire EV of luminance. The ETTR theory is that you can improve that detail by increasing exposure (if possible, of course.) What I'm saying is that yes, the digital value from the photosite is limited to three values (1, 2, and 3) but the final values for the pixel will be calculated at a greater bit depth, and will obtain greater discrimination in detail through the influence of surrounding photosites. So it's not as bad as it seems.

[Colin Southern]

But that's not related to the "levels" issue. The levels issue is related to the shadow end, where fewer and fewer bits are used to describe the luminance level. So on a 12-bit camera, when you get 10-stops down from the max signal level you only have 3 values to describe an entire EV of luminance. The ETTR theory is that you can improve that detail by increasing exposure (if possible, of course.) What I'm saying is that yes, the digital value from the photosite is limited to three values (1, 2, and 3) but the final values for the pixel will be calculated at a greater bit depth, and will obtain greater discrimination in detail through the influence of surrounding photosites. So it's not as bad as it seems.

On a 12 bit camera - with a non-ETTL exposure - shadow detail is likely to be in the 4 stops above the noise floor area (so 16 levels) - but usually having more levels isn't going to make an image more appealing because increased resolution of that type of detail just isn't noticeable in a normal photo. If one is having to reveal that kind of detail to the point where the blocking due to lack of bits is obvious then (a) the whole photo was probably significantly under-exposed to start with (assuming a normal reflective scene), or (b) it's a scene that needs HDR handling in which case ETTR may be sufficient or it may not (in which case we're into GND or bracketed exposure territory).

It's a bit like chrome plating the bolts inside an engine; in theory it makes for a higher-quality product, but in practice, nobody can see it.

[pnodrog]

But that's not related to the "levels" issue. The levels issue is related to the shadow end, where fewer and fewer bits are used to describe the luminance level. So on a 12-bit camera, when you get 10-stops down from the max signal level you only have 3 values to describe an entire EV of luminance. The ETTR theory is that you can improve that detail by increasing exposure (if possible, of course.) What I'm saying is that yes, the digital value from the photosite is limited to three values (1, 2, and 3) but the final values for the pixel will be calculated at a greater bit depth, and will obtain greater discrimination in detail through the influence of surrounding photosites. So it's not as bad as it seems.

This assumes the sensor output is converted completely linearly during the AD conversion process. One technique I am aware of is that the initial conversion starts using a higher reference voltage on the ADC giving the shadows a steeper slope so instead of say for a normal linear conversion value of 3 you may end up with say 8 as the value and it quickly increases over the next few stops. Once the conversion of the photosite (pixel) reaches a predetermined level which the application designer has nominated (mid tones and/or highlights) the ADC reference voltage is switched to the normal full scale voltage and the conversion slope for the rest of the conversion is effectively reduced. These sensors I understand are used in some security surveillance cameras.

I have no idea about the conversion methods used by Canon, Sony or requested? by Nikon and they are unlikely to tell me. They may be multiple slope or continuously variable slope but I would be surprised if they are still using a direct linear conversion and if they are I would not bank on it continuing. Some (most?) of the ETTR theory is based on the assumption that the conversion is linear.

The best way is still the old way - get to know how your particular camera sensor (used to be film) behaves and adapt to produce the results you want. What is the best approach with your present equipment is unlikely to be the best with a future camera. Yep - practice not theory.

[Graystar]

I know that every camera is different, but here's a progression of RAW values and matching 8-bit sRGB values from my Nikon D90. I took 10 images, the first representing the maximum unclipped signal level for the green channel. Each subsequent image had exposure reduced by one EV via shutter speed. The RAW values came from RawDigger, and the sRGB values come from Raw Therapee with a Neutral profile (which basically means white balance only.)

RAW - sRGB 8-bit (difference from row below)
3830 - 255 (+67)
1909 - 188 (+52)
950 - 136 (+37)
475 - 99 (+29)
237 - 70 (+21)
119 - 49 (+15)
60 - 34 (+13)
30 - 21 (+8)
16 - 13 (+5)
9 - 8

There are a few interesting things. If you start at 3830 and just keep dividing by 2, the RAW data follows a halving of values fairly consistently. The sRGB values don't appear to follow any progression, which is due to the gamma factor being thrown in there. Another thing to note is that even 10 EV down, the RAW data contains more levels than sRGB 8-bit data, and does so through the range. Things would be different with 16-bit data, but alas, our monitors are still 8-bit devices, so that's the limit we have to live with visually. But I believe the situation would be the same even with 10-bit data and 10-bit displays...and 10-bit displays are known to have smooth gray gradients. Just another reason why I don't worry about levels.

I also find it interesting that sRGB 99 is three stops down from the max, which is exactly right...it's 12.5% of the max value and that calculates out to sRGB 99. On average I get sRGB 100 when I spot-meter a neutral reference (about 12.8% reflectance.) If I were to use a Kodak gray card to set exposure, and I follow the instructions with the card that says to increase exposure by 1/2 EV, I'd get the max RAW signal from bright whites. I know that Nikon D7000 and D800 bodies are the same way.

[pnodrog]

Interesting that the D90 is similar to the D800. I used to assume the RAW file was the data directly from the ADC output but the more little pieces of information I come across and the more I think about it the less certain I become. Certainly there is hardware available that can subtract/substitute the dead photosites and remove some noise from the output but I suspect the RAW files are not as RAW as we assume - after all, the manufactures don't call it RAW and they adopt differing formats and they need to give the camera software engineers something interesting to play with. To understand it at the level I would like to I would have to spend so much time researching that I would have very little time to use my camera.

[revi]

But that's not related to the "levels" issue. The levels issue is related to the shadow end, where fewer and fewer bits are used to describe the luminance level. So on a 12-bit camera, when you get 10-stops down from the max signal level you only have 3 values to describe an entire EV of luminance. The ETTR theory is that you can improve that detail by increasing exposure (if possible, of course.) What I'm saying is that yes, the digital value from the photosite is limited to three values (1, 2, and 3) but the final values for the pixel will be calculated at a greater bit depth, and will obtain greater discrimination in detail through the influence of surrounding photosites. So it's not as bad as it seems.

I see now what you mean.

Sorry to say that I cannot agree, though.
First, if you are that low in the curve, your noise is getting very important, and your values will be precise to ± 1 (at best, so a registered value of 1 could just as well be 0 or 2 in reality). Then, yes interpolation will give a bit more information, but if you are in an area of low signal, the neighbouring pictures will be close to the same values, with the same uncertainty/error. So you might get values expressed in more bits, but your information content will not have increased above the 2-bit level. Which will be visible (either no detail, or 'smeared' noise blobs).

Note that the non-linear A/D conversion mentioned might help a bit, but we'll still be dealing with the inherent noise. And I don't know what the inherent noise level of a sensor is (it's probably related to the number of photons captured, plus a thermal component).

[xpatUSA]

Interesting that the D90 is similar to the D800. I used to assume the RAW file was the data directly from the ADC output but the more little pieces of information I come across and the more I think about it the less certain I become.

Nikon uses lossy compression on some (e.g. D50), not on others (e.g. D1), and selectable to either on yet others.

12 bit raw image data is compressed by using a lossy encoding followed by non adaptive Huffman compression.
During encoding the 4096 possible distinct values from the 12 bit Analog to Digital Converter (ADC) are reduced to either 567 or 683 depending on the camera model.

Yep, the much-vaunted .NEF file from many Nikon cameras has had lossy compression applied and arrives at the RAW converter with only 683 levels - goodbye, 12-bit resolution, ho ho. Did you know that?

You can read more and see the curve here:

http://kronometric.org/phot/NEF_Compression.doc

The ETTR plot is thickened somewhat by Sigma who, of course, just have to be different. My camera has "saturation" values embedded in it's firmware to match the installed sensor, but they are not true "well full" values. Reason is that the sensor is not the world's most linear from what I've read. So, in the RAW (X3F) file, values of about 6-7,000 for "saturation" are found but, if you really saturate the poor old sensor, much higher values of around 9,000+ can be seen. This makes the area from 6,000 to 9,000 or so a true adventure playground for ETTR Foveonistas ;-)

BTW, please don't inform me that 6,000 is more than 12-bits - Sigma's conversion algorithms are not that simplistic.

[pnodrog]

My little old Nikon D200 gives me an uncompressed NEF file and the size varies slightly but is always large enough to cater for 12 bits for each pixel. I have tried using the compressed NEF and could not really see any difference in the image quality. When using photoshop at 8 bits per channel the effects of the compression will be undetectable.

On the bright side these types of post help give me odd little insights into the technology - the draw back is it takes up time that maybe better used.