ETTR with strong yellow

[bjanes]

Since reading Michael's seminal article on exposure to the right, I have been using this technique with good results. Placing the highlights just short of clipping can be difficult and Michael suggested using the histogram. But which histogram? Most cameras have a luminance histogram whose results are heavily weighted towards green, and many newer cameras have RGB histograms which are useful to show clipping in the RGB channels. At first glance, with strong colors, one might assume that the RGB histogram is preferable. However, this does not take white balance into consideration.

Here are histograms from the Nikon D3 I got when taking a snapshot of a yellow flower with the camera set to aRGB and auto WB with daylight illumination. The first exposure was obtained from the cameras matrix metering and showed strong clipping in the green and red channels. The middle and bottom views show decreasing exposure (0.33 EV decrements). It was not possible to avoid clipping of the red channel with any reasonable decrease in exposure, so I stopped when the green channel was no longer blown.

[attachment=7215:attachment]

One can view the raw histograms using Rawanalyze. The red and green channels that were blown in the camera histogram of the original exposure are actually intact in the raw file. According to DCRaw, the multipliers needed for white balance are 1.699 and 1.512 for the red and blue channels respectively. The green multiplier is left at 1.0. Clipping occurs with white balance when the multipliers are applied.

[attachment=7216:attachment]

With the original exposure, green clipping can be eliminated in aRGB with a -0.9 EV exposure compensation, but it is not possible to remove red clipping by decreasing the exposure.

[attachment=7217:attachment]

The red simply won't fit into aRGB, but with ProPhotoRGB, the red clipping can be nearly eliminated with -1.45 EV exposure compensation.

[attachment=7218:attachment]

In this case, the luminance histogram is more helpful than the RGB histogram. From the white balance coefficients, the red channel has 0.76 f/stops of headroom and the blue channel 0.6 stops. One could estimate this headroom on the camera RGB histogram or else load a UniWB into the camera. With the UniWB, the WB multipliers are all 1.0 and give a better estimate of the actual status of the raw channels. However, one is still in aRGB since this camera does not allow ProPhotoRGB as a color space. Another approach would be to place a magenta filter in front of the lens to hold back some of the green and better balance the raw channels.

The extra 0.67 EV exposure of the last shot should give a better signal:noise and better dynamic range, but I was unable to see any significant difference in the actual results. The exercise does show the advantage of using PhotoProRGB.

Bill

[Jonathan Wienke]

RGB histograms are only useful when the camera's internal white balance is set so that each channel has a unity white balance multiplier, AKA UniWB. Otherwise the camera white balance setting will nullify the advantage an RGB histogram offers--seeing when a single RAW channel is blown. Search the forums for UniWB and you'll find a rather lengthy thread on the subject and the solution to your problem.

[bjanes]

RGB histograms are only useful when the camera's internal white balance is set so that each channel has a unity white balance multiplier, AKA UniWB. Otherwise the camera white balance setting will nullify the advantage an RGB histogram offers--seeing when a single RAW channel is blown. Search the forums for UniWB and you'll find a rather lengthy thread on the subject and the solution to your problem.
[a href=\"index.php?act=findpost&pid=203893\"][{POST_SNAPBACK}][/a]

The corollary of your statement is that RGB histograms are rarely useful, since few photographers load UniWB into their cameras. UniWB is relatively simple with Nikon SLRs, but more difficult with Canon.

You must not have read my post carefully, since I did mention UniWB. Actually, if you know the WB multipliers, you can interpret RGB histograms up to the point of clipping. Even with UniWB, the histogram is still based on the working space in the camera, and saturation clipping can easily occur if this space is Adobe RGB.

Bill

[Jonathan Wienke]

The corollary of your statement is that RGB histograms are rarely useful, since few photographers load UniWB into their cameras. UniWB is relatively simple with Nikon SLRs, but more difficult with Canon.

Agreed, on all points.

You must not have read my post carefully, since I did mention UniWB. Actually, if you know the WB multipliers, you can interpret RGB histograms up to the point of clipping.

Sorry, I saw the part where you mentioned having the camera on auto white balance and missed the UniWB reference. Auto WB is the worst possible setting for histogram-to-RAW consistency, since the relationship between RAW clipping and the histogram is changing to some extent with every shot. Using a non-UniWB setting requires one to interpret the histogram more than UniWB. It can be done, but adds complexity to correctly reading the histogram.

Using a small color space in-camera (or one that fits completely within the actual sensor gamut) is not necessarily a bad thing, since saturated colors will show clipped color channels in the histogram before the RAW data clips.

[Panopeeper]

if you know the WB multipliers, you can interpret RGB histograms up to the point of clipping

I doubt that you can do that with anl accuracy, which makes it useful. The histogram is based on the non-linear transformation, which can not be turned off. This is the biggest factor for the in-camera raw histogram not being representative for the raw data.

Luckily, this is not so bad, because clipping can be judged this way; I find the method with UniWB perfect. However, the method is hardly suitable to determine, how far a channel is from clipping (due to the non-linear transformation). The closer the channel is to the right edge, the smaller the visible change caused by an increment of the exposure. Thus, a multiplication by a fix number "expands" or "shrinks" the visible histogram by a varying degree, depending on its actual "size".

[bjanes]

Sorry, I saw the part where you mentioned having the camera on auto white balance and missed the UniWB reference. Auto WB is the worst possible setting for histogram-to-RAW consistency, since the relationship between RAW clipping and the histogram is changing to some extent with every shot. Using a non-UniWB setting requires one to interpret the histogram more than UniWB. It can be done, but adds complexity to correctly reading the histogram.
[{POST_SNAPBACK}][/a]

Your point about autoWB is well taken. Since I shoot in raw, I often leave the WB at auto in daylight work, since the results in the preview are usually satisfactory and the WB can be set later.

With strong colors in only one primary the auto WB may not work well, especially if there are no neutrals in the field of view. I don't know the details of how auito WB works with the D3, but Sean McHugh implies that with some cameras the presence of a neutral white in the image is helpful:

[a href=\"http://www.cambridgeincolour.com/tutorials/white-balance.htm]Cambridge in Color [/url]

With the yellow flower, it would have been best to include a view with a WhiBal card or use a preset WB. Lacking that, I suppose a daylight WB would be better.

Bill

[bjanes]

I doubt that you can do that with anl accuracy, which makes it useful. The histogram is based on the non-linear transformation, which can not be turned off. This is the biggest factor for the in-camera raw histogram not being representative for the raw data.

Luckily, this is not so bad, because clipping can be judged this way; I find the method with UniWB perfect. However, the method is hardly suitable to determine, how far a channel is from clipping (due to the non-linear transformation). The closer the channel is to the right edge, the smaller the visible change caused by an increment of the exposure. Thus, a multiplication by a fix number "expands" or "shrinks" the visible histogram by a varying degree, depending on its actual "size".
[a href=\"index.php?act=findpost&pid=204005\"][{POST_SNAPBACK}][/a]

The gamma encoding in the rendered file does complicate things, as does the tone curve used in the camera, but these difficulties can be overcome with observational data. For example, I photographed a Stouffer step wedge with the Nikon D3 with the Picture Control set to Standard (contrast = 0, brightness = 0, saturation = 0). I then rendered the file into aRGB with Nikon Capture NX, which produces results very similar to those produced by the in camera JPEG from which the camera RGB histograms are derived.

I then plotted the raw pixel values of the green1 channel (obtained from your Rawanalyze) against the value of the green channel in aRGB obtained in NX. A similar exercise could be done for the blue and red channels.

Most readers will be more familiar with the linear plot, but the LogLog plot demonstrates the data to better advantage.

[attachment=7241:attachment]

[attachment=7242:attachment]

Bill

[Panopeeper]

Bill,

I do not doubt the validity of the data you presented. However, I do dpubt, that you can make use of this, when shooting. Let's say it is daylight and the respective WB coefficients are 1.58, 1.0, 1.35 (I just made up these numbers). Now, you need to reverse the aRGB or sRGB transformation (the "gamma"), divide the result by 1.58, multiply it by 1.62 (2/3 steps increase of the exposure, as an example), and transform it again; now you "know", how the raw histogram would look like. LOL

I put up with the fact, that the tumbnails and the embedded JPEG images are only for finding the right image; all are green and greener (as the consequence of the neutral WB). I put up with the fact, that I can be 1/3 EV or 2/3 EV (or more) away from the optimal exposure after adjusting it, but I know, that when the histogram is *almost* at the right edge, then I am close to the ideal exposure, and, more importantly, I know that the blinking on the LCD indicates *very* reliable if and how much clipping occured.

My pain is only, that my sucker camera does not support quintett exposure bracketing.

[bjanes]

Bill,

I do not doubt the validity of the data you presented. However, I do dpubt, that you can make use of this, when shooting. Let's say it is daylight and the respective WB coefficients are 1.58, 1.0, 1.35 (I just made up these numbers). Now, you need to reverse the aRGB or sRGB transformation (the "gamma"), divide the result by 1.58, multiply it by 1.62 (2/3 steps increase of the exposure, as an example), and transform it again; now you "know", how the raw histogram would look like. LOL
[a href=\"index.php?act=findpost&pid=204099\"][{POST_SNAPBACK}][/a]

We are interested less in the shape of the raw histogram than the right end where clipping is indicated. If we assume daylight WB with red and green multipliers of   1.699 and 1.512 for the red and blue channels, they would clip at raw values of about 9400 and 10600 respectively. With the default camera tone curve mentioned previously and UniWB, these values would be rendered at approximately 236 and 235 in aRGB, as shown on this graph which represents the camera tone curve, which consists mainly of the gamma "correction" but also includes the S curve of the default rendering:

[attachment=7246:attachment]

If you applied white balance, the red and blue raw channels would be raised to about 16000, and all channels would clip at about 255 in the rendered file. Since we are on the flat portion of the tone curve, large differences in raw values result in much smaller differences in the rendered pixel values. You might not be able to see much difference between the right ends of the UniWB and normal histograms.

My approach with normal WB RGB histograms would be to look at the green channel first. Since the green multiplier is 1.0, any clipping here indicates clipping in the raw green channel and exposure must be reduced. In my previous example, the RGB of the metered exposure is shown:

[attachment=7247:attachment]

Since the green channel is blown, the exposure must be reduced. Reducing exposure by 0.66 EV results in this camera histogram:

[attachment=7248:attachment]

The red channel still shows clipping, but the green channel is fine. Since the red channel of the raw file is 0.76 stops to the left of the green, if we reduce exposure by another 0.67 EV, we would have to be at or below clipping in this channel and we could stop. At this point, clipping in the red histogram would most likely represent saturation clipping in aRGB.  The blue channel is 0.6 EV to the left of green, and similar considerations would apply.

Therefore, with normal daylight white balance I would recommend bracketing downward 0.67 stops (two clicks on the camera exposure controls) from green saturation if clipping is evident in the red or blue channels. Any minor residual clipping could be handled with highlight recovery. To be safe, one could bracket downward by a full f/stop.

I put up with the fact, that the tumbnails and the embedded JPEG images are only for finding the right image; all are green and greener (as the consequence of the neutral WB). I put up with the fact, that I can be 1/3 EV or 2/3 EV (or more) away from the optimal exposure after adjusting it, but I know, that when the histogram is *almost* at the right edge, then I am close to the ideal exposure, and, more importantly, I know that the blinking on the LCD indicates *very* reliable if and how much clipping occured.

My pain is only, that my sucker camera does not support quintett exposure bracketing.
[a href=\"index.php?act=findpost&pid=204099\"][{POST_SNAPBACK}][/a]

This is in agreement with my observations.

Bill