Histogram: is it good to follow or not

PhotosGuy wrote in post #4468788
Maybe because some highlights aren't important & saving them would just underexpose the important areas of the shot? Which is why I usually use this to keep the important whites white: Need an exposure crutch?

Perfect, then you would just spot meter over the highest lights you are really interested to preserve, and let the camera allocate them to the right of the histogram, blowing any pixels with higher luminance than those you spot metered. It's the photographers's decision whether to preserve the entire scene highlights, or just some of them in order not to underexpose the main subject.

Besides all that, we must know that ALL shots that we consider perfectly exposed to the right (i.e. where histogram is to the right but highlights have been preserved) have actually some pixels blown, it's simply the number is not high enough to be noticeable.

In my example pic from the Namibia desert more than 5000 pixels had the G channel blown: Burnt {R;G;B} levels: {90;5137;163} -> {0%;0,1%;0%}
but they didn't affect at all the gross part of the information which started 1/4 f-stop later. Looking at the histogram it's clear why: 5000 pixels in a 8Mpx camera are negligible. And this has to be this way, otherwhise to preserve strictly all highlights pixel by pixel, we would ALWAYS get a terribly underexposed image.

In a ETTR exposure mode the % of pixels that are allowed to be blown would be a fantastic user parameter. We would set for instance 0,1%, and the camera would calculate exposure to force that 0,1% of the pixels, not more not less, would be blown. For instance DCRAW has a -b parameter which sets exactly this when developing any RAW file, with dcraw -b 0.1 the program calculates the exposure correction needed to force 0.1% of the image's pixels to be blown (white), so overall exposure in the image is adjusted properly if the user accidentally underexposed the shot. I claim it would be great to have this in-camera so we let the camera make sure the that the RAW file contains the highest possible quality of information.

_GUI_ wrote in post #4469000
...
In a ETTR exposure mode the % of pixels that are allowed to be blown would be a fantastic user parameter. We would set for instance 0,1%, and the camera would calculate exposure to force that 0,1% of the pixels, not more not less, would be blown. For instance DCRAW has a -b parameter which sets exactly this when developing any RAW file, with dcraw -b 0.1 the program calculates the exposure correction needed to force 0.1% of the image's pixels to be blown (white), so overall exposure in the image is adjusted properly if the user accidentally underexposed the shot. I claim it would be great to have this in-camera so we let the camera make sure the that the RAW file contains the highest possible quality of information.

My own experience leads me to a slightly different algorithm, as follows:
Specular reflections are generally head & shoulders out in front of non-specular reflections and should receive special treatment. How do we recognize specular reflections? They tend to have a low color variance, and they are much higher in value than the normal highlights. Therefore, it should be possible to create an algorithm that permits those areas judged as being specular to blow out, unrestrained.

Robert_Lay wrote in post #4469296
My own experience leads me to a slightly different algorithm, as follows:
Specular reflections are generally head & shoulders out in front of non-specular reflections and should receive special treatment. How do we recognize specular reflections? They tend to have a low color variance, and they are much higher in value than the normal highlights. Therefore, it should be possible to create an algorithm that permits those areas judged as being specular to blow out, unrestrained.

Yes, this kind of improvements over the basic behaviour would be OK. For instance detecting in the scene representative clusters of information with luminance empty gaps between them in the histogram, and setting some rule such as: "any cluster of pixels that contains less than 5% of total pixels and is several (4 or 5 or wathever) f-stops above the rest of the image, is likely to be a specular reflection so it has to be left out of any metering consideration and we will allow it to appear blown in the final RAW file.

I think there is room for a lot of improvement in digital cameras regarding light metering systems, and definitively real time pre-analysis of the scene will give us a lot of control when camera vendors focus on this area of research. I have expectances since cheap compact cameras can calculate real-time histograms, it cannot be that difficult to implement this kind of things.

_GUI_ wrote in post #4469450
Yes, this kind of improvements over the basic behaviour would be OK. For instance detecting in the scene representative clusters of information with luminance empty gaps between them in the histogram, and setting some rule such as: "any cluster of pixels that contains less than 5% of total pixels and is several (4 or 5 or wathever) f-stops above the rest of the image, is likely to be a specular reflection so it has to be left out of any metering consideration and we will allow it to appear blown in the final RAW file.

I think there is room for a lot of improvement in digital cameras regarding light metering systems, and definitively real time pre-analysis of the scene will give us a lot of control when camera vendors focus on this area of research. I have expectances since cheap compact cameras can calculate real-time histograms, it cannot be that difficult to implement this kind of things.

I agree with your comments, and those of Robert, but there is a practical problem which the marketers must consider:

How many users or buyers of Canon cameras would understand and be able to use this feature? And what if they set it incorrectly?

My guess is maybe 0.10 percent - about the percentage of pixels you want to allow to be blown. That's 1/1000 for the mathematicall challenged.

About specular highlights mentioned by Robert - they seldom if ever have any (or much) real colour rendition do they? From my perception, the reflection from a rose leaf is pretty much the same colour as that from a car windshield.

Don't they actually have no color variance as they're 255,255,255?
(Oops! Glenn got there first., But I typed it & will stand behind it... or behind him)

PhotosGuy wrote in post #4470280
Don't they actually have no color variance as they're 255,255,255?
(Oops! Glenn got there first., But I typed it & will stand behind it... or behind him)

According to my own, personal model of the RAW image file and how clipping occurs, there is "hard clipping" and there is "soft clipping".

Hard Clipping: All 3 channels become 255 in the camera at the time the picture is taken.
Soft Clipping: Any one or more of the channels reaches 255 during RAW processing.

The short version of the story is that soft clipped pixels are recoverable in Adobe Camera RAW, while hard clipped pixels are not recoverable.

My theory on this topic is explained in greater detail at:
*************RAW******​*********
Tutorial on RAW Processing as downloadable PDF:
http://www.zaffora.com​/W9DMK/RAWProcessing.p​df

Quick Tour of RAW Processing as downloadable PDF:
http://www.zaffora.com …ckTourOfRAWProc​essing.pdf
[The latter has the latest material]
***************

Back to the question of the pixel in a specular reflection - I believe that it can be either hard or soft clipped. It's only a question of how bright it is to the sensor.

Robert_Lay wrote in post #4472449
According to my own, personal model of the RAW image file and how clipping occurs, there is "hard clipping" and there is "soft clipping".

Hard Clipping: All 3 channels become 255 in the camera at the time the picture is taken.

Wouldn't this be 4096 for a raw file (12-bit)? I'm not nitpicking, it's actually an important point.

Peter Pawinski wrote in post #4472798
Wouldn't this be 4096 for a raw file (12-bit)? I'm not nitpicking, it's actually an important point.

Or 16,380 in 14 bit bit, but in the context of this discussion it is not an important point. The important point is that the output has hit the top level, however you quantify it.

Glenn NK wrote in post #4470060
How many users or buyers of Canon cameras would understand and be able to use this feature? And what if they set it incorrectly?

My guess is maybe 0.10 percent - about the percentage of pixels you want to allow to be blown. That's 1/1000 for the mathematicall challenged.

About specular highlights mentioned by Robert - they seldom if ever have any (or much) real colour rendition do they? From my perception, the reflection from a rose leaf is pretty much the same colour as that from a car windshield.

I totally agree. It's a matter of building a camera that satisfies most target customers. And advanced "strange" features like these would be understood and appreciated by few crazy people like us. So, let's try everyone get crazy!

Robert, about specular lights, you must think that if we assume they are some f-stops ahead of the rest of pixels of the scene, it is very difficult that only 1 or 2 channels are blown in them unless the overall image is terribly underexposed. The usual case will be that the three channels will be blown by far at the same having thus (255,255,255) values, "hard-clipping" in your article's terminology (BTW very interesting). I cannot think of any natural source of light having R channel saturated in the sensor, and the G or B channels some 3 or 4 f-stops below (i.e. at a fraction of 12%-6% the linear intensity of the R channel). Maybe a red LASER light source could achieve this, but not a natural shine or reflection.

Maybe not "3 or 4 f-stops below", but I've seen shots were the WB wasn't set & one channel is blown with the others not. Makes it hard to WB the raw file afterwards. I think Curtis had a thread on that somewhere.

EDIT
Found it: How NOT to expose to the right

And, Rene found this: Restore Those Clipped Channels

PhotosGuy wrote in post #4477131
Maybe not "3 or 4 f-stops below", but I've seen shots were the WB wasn't set & one channel is blown with the others not. Makes it hard to WB the raw file afterwards. I think Curtis had a thread on that somewhere.

Of course, that happens many times. But we were discussing about specular lights: reflections on glasses or metallic elements that usually appear completely white. Or even direct light sources (the sun, light bulbs,...). Those shiny parts are some f-stops above the rest of the scene, so unless you underexpose severely the entire image, their content will appear completely white, not partially blown.

tzalman wrote in post #4474268
Or 16,380 in 14 bit bit, but in the context of this discussion it is not an important point. The important point is that the output has hit the top level, however you quantify it.

Fine. It's an important point to me conceptually, but I understand, I guess, if it isn't to others. RAW didn't make sense to me until it was shown to me how what a JPEG may show as 255 as blinking highlights on the LCD preview may not actually be clipped data in RAW, because there are a range of values that resolve to 255 when expressed as an in-camera JPEG (or PP develop settings), but that DO contain detail in the raw capture.

Just because highlights clip on the back of the camera (show 255) does not mean raw data is necessarily clipped (value of 4096/16,380). That's why I consider it an important distinction. Personally, it's confused me in the past and understanding the two different sets of numbers helps keep it clear in my head.

Peter Pawinski wrote in post #4472798
Wouldn't this be 4096 for a raw file (12-bit)? I'm not nitpicking, it's actually an important point.

I don't think so.
Wouldn't it make more sense that in the 12 bit domain there would be a range of many values that would all downsample to the value 255?

For example, would it not be logical for everything from 2028 to 2036 to all downsample to 127?

I think it will be simpler if we were all to stay with the 8 bit world, since there is a lot of fog lying between the 12 bit RAW world and the 8-bit JPG world.

Robert_Lay wrote in post #4479448
I don't think so.
Wouldn't it make more sense that in the 12 bit domain there would be a range of many values that would all downsample to the value 255?

For example, would it not be logical for everything from 2028 to 2036 to all downsample to 127?

Yes, that's exactly my point, unless I'm misunderstanding something. When you have a spread of values that all downsample to 255, you could have discrete details where there doesn't appear to be discrete details, no? Isn't this how what appear to be blown highlights on a basic raw import get translated into detail? So if you have a file where the highlights are at 4095, 4092, 4090 -- when they get downsampled they all become 255. However, you can manipulate that data to extract highlight information which would otherwise appear lost.

From a raw capture standpoint, data is only lost on the input end when all three channels clip at 4095. It doesn't matter how many of those values map to 255 in an 8-bit RGB colorspace, because you can remap the RAW as you wish (which is what you do with ETTR, anyway).

I have a feeling we're all on the same page here, but talking about different things. What I'm saying is that what maps to 255 on the back of your LCD, which shows you a pre-defined JPG rendering of a RAW file, is not the same as hitting 4095 in your RAW file precisely because many values can map to 255, and none of them might be 4095. I don't know exactly how many values map to 255, but it's a lot more values than map to 127 and a heck of a lot more values than map to 1, precisely because the relationship between raw values and RGB values are not linear.

Peter Pawinski wrote in post #4479798
Yes, that's exactly my point, unless I'm misunderstanding something. When you have a spread of values that all downsample to 255, you could have discrete details where there doesn't appear to be discrete details, no? Isn't this how what appear to be blown highlights on a basic raw import get translated into detail? So if you have a file where the highlights are at 4095, 4092, 4090 -- when they get downsampled they all become 255. However, you can manipulate that data to extract highlight information which would otherwise appear lost.

From a raw capture standpoint, data is only lost on the input end when all three channels clip at 4095. It doesn't matter how many of those values map to 255 in an 8-bit RGB colorspace, because you can remap the RAW as you wish (which is what you do with ETTR, anyway).

I have a feeling we're all on the same page here, but talking about different things. What I'm saying is that what maps to 255 on the back of your LCD, which shows you a pre-defined JPG rendering of a RAW file, is not the same as hitting 4095 in your RAW file precisely because many values can map to 255, and none of them might be 4095. I don't know exactly how many values map to 255, but it's a lot more values than map to 127 and a heck of a lot more values than map to 1, precisely because the relationship between raw values and RGB values are not linear.

One point that you might have overlooked is that 4096 discrete tonal values in the RAW virtual image allow for 12 f-stops of dynamic range, while 256 discrete values in the 8-bit JPG only allow for 8 f-stops of dynamic range. Therefore, when the RAW file is processed to 8-bit JPG are you giving up detail - or are you giving up dynamic range?

I would not even attempt to explain the Bayer de-mosaicing algrorithm, but I do think that there is so much happening in that process that it may be beyond our ability to get a grip on all of what takes place.