Optimal bracketing technique

[feppe]

Long-time lurker, first-time poster  This turned out a bit long.

I shoot a lot of very long exposures (30+ seconds) while doing nighttime cityscapes. The problem with the subject matter is that the dynamic range is well beyond the (theoretical) 8-stop range of my camera (Canon 30D): getting detail in both shadows and highlights is impossible with just one exposure.

The usual response is to bracket. This works well but bracketing the usual 3 or 5 shots around the metered exposure becomes unwieldy fast with nighttime scenes. A typical metered nighttime cityscape requires 30-90 sec exposure and excessive bracketing can mean tens of minutes of extra time shooting with several scenes and especially while doing stitched panoramas.

For this reason I've come up with a technique to minimize bracketing needs to 2 shots. The goal is to get 2 exposures which will be blended in post for a 16-bit extended dynamic range photo. 2 shots should theoretically be able to cover up to 16 stops with my 30D with a 8-stop dynamic range according to dpreview.com. As most scenes require less than 16 stops there is plenty of room for "overlap." Further, it should be apparent that 3-shot bracketing is unnecessary most of the time.

So, here we go:

Step 1: Use a spot meter to meter the darkest part of the scene where you want detail in the final blended exposure.
Step 2: Expose a shot X stops above the metered exposure. X is the exposure latitude of the sensor you're using. For Canon 30D it is 5 stops below and 3 steps above metered according to dpreview.com. I'd use 4 (X) and 2 (Y in Step 4 below), respectively, to be on the safe side. So, if I meter 30sec at f/8 I'd shoot ~8min at f/8.
Step 3: Spot meter the lightest part of the scene where you want detail.
Step 4: Expose one exposure Y stops below the metered exposure. If I meter 10sec at f/8 I'd shoot 2.5sec at f/8.

This would yield two shots with a dynamic range of ~8 stops (4 stops below the dark exposure, 2 above the light and ~2 in "between" the exposures) after blending. This is on par with the theoretical 8-stop range of 30D but should have significantly less noise in the shadows than one un-bracketed exposure shot as metered.

This technique is meant to bring the darkest and lightest parts with detail just in the range of the sensor while maximizing the final stop range and contrast, and minimizing the number of exposures needed. It should be easy to cover a scene with a range of, say, 11-stops with this technique. Blending these exposures to yield a 16-bit file with those 11 stops would be the next step.

Now, I haven't tried this in the wild yet, and my theory might be irredeemably faulty. Therefore I would like to hear you guys - who know much more about this than I do - to shoot this technique down.

[wood]

Hello,

Did you try shoot on RAW??

More Flexibility

When you shoot in JPEG or TIFF, the camera will apply its characteristic curve to the RAW data and this may clip highlight or shadow detail by increasing contrast (which creates the perception of larger dynamic range).
This curve depends on both the camera design and the user setting.
However, if you process the RAW data on your PC, you are in control of which curve you apply the RAW data.

RAW can Replace Bracketing

If the sensor has a dynamic range of more than 256 levels per channel, saving the data in JPEG or 8 bits per channel TIFF format will clip the dynamic range.
In RAW mode you have access to the complete dynamic range as captured by the sensor.
You may for instance prefer to extract more shadow or highlight detail than the camera would have done by applying the characteristic curve and saving it in 8 bits per channel JPEG or TIFF.
The ability to extract detail is equivalent to having exposure flexibility.
Indeed, if you do not have the option of RAW(but you have with your 30D), you would have to resort to bracketing in JPEG or TIFF to capture the same dynamic range.
But it will take more time to shoot several bracketed JPEG images and their combined size will be the same as one RAW file, even if you use a tripod, there could be minor misalignments between the images, making it difficult to join them afterwards into one image.
So instead of bracketing your images on the scene by shooting multiple JPEG or TIFF images, you can shoot one or two RAW image instead (with +2,3 and -2,3 stops, to help you with  highlight and shadows)  and extract the desired detail on your computer.
You could for instance extract one image with shadow detail and one image with highlight detail by adjusting the Exposure slider in PhotoshopCS.    
Combining these images is easier because both layers will be perfectly aligned.

Wood

[jani]

Did you try shoot on RAW??

The problem the original poster is describing is also present when you shoot in raw. The dynamic range of the camera is still limited; don't buy into the hype that raw solves that automagically.

However, if you process the RAW data on your PC, you are in control of which curve you apply the RAW data.

RAW can Replace Bracketing

(...)

You could for instance extract one image with shadow detail and one image with highlight detail by adjusting the Exposure slider in PhotoshopCS.   
Combining these images is easier because both layers will be perfectly aligned.

I have yet to find any convincing evidence that two raw conversions of the same file are better than one, for extracting the maximum dynamic range.

If you work well enough with the highlights, midtones and shadows in your raw conversion tool (and the tool lets you do it!), then that extra conversion should be unnecessary, IMHO.

But I suppose it depends on your workflow.


As for the original poster's idea, I think it's interesting, but is it usable in the field? Perhaps it's a matter of personal taste and styles. With my 20D, I've used the center-weighted average meter to guesstimate exposure for the dark areas, then guesstimate exposure for the bright areas, set the exposure in the middle, and used exposure compensation of +/- 2 Ev to get approximately what I want. The 20D in raw mode may have a "10 stop" dynamic range according to DPreview, but due to noise concerns when bracketing, it's nice to pretend that it's a couple of stops less, at least. Or so I think.

Unfortunately, my success with finding a compelling subject for compressed dynamic range imaging (referred to by others as "HDR", or high dynamic range imaging) has been abysmally low, so I've not used it a lot. It seems to be mostly a toy, but it can have its use on special subjects.

[feppe]

I only shoot RAW and edit in 16-bit ProPhoto RGB until ready to print. As RAW out of 30D is 12-bit and I use 16-bit TIFF files out of the converter, I also don't see how two RAW conversions would benefit. And as I'm mostly concerned about noise in dark areas and increasing contrast (so I have more room to tweak it down later) pushing the shadows to the sensor's limits in C1 or whatever is counter-productive, IMO. Further, I've never had any problems with alignment with pictures as I use a sturdy tripod and there are programs that do it automatically in case there's a mishap.

I'm not talking about HDR - the pics one can see on flickr under HDR tag for example. I've yet to see a HDR photo I like as I think they look unreal, painterly and "dirty." I'm talking about just general digital blending to increase dynamic range, like described in the Luminous Landscape article on digital blending. I've had success with it but am looking for a bit more optimized solution so I don't end up wasting precious time.

Bracketing two stops over and below would work and I tend to use that. But as I said, it becomes unwieldy especially when shooting stitched panoramas. Cutting it down to 2 would save 1/3 in shooting time, which can be critical during sunsets, for example.

30D in RAW has a 8-stop dynamic range according to dpr - not 10 - and I'm sceptical about that as Jan is. That's why I'm working with the assumption that it is in fact 6 stops for the same reason Jan does.

[jani]

30D in RAW has a 8-stop dynamic range according to dpr - not 10 - and I'm sceptical about that as Jan is. That's why I'm working with the assumption that it is in fact 6 stops for the same reason Jan does.

The best we could achieve (with some pretty extreme ACR settings) was just over 10 stops total dynamic range, and more importantly about a stop more highlight range (although with no guarantee of color accuracy).

(page 21)

[Jonathan Ratzlaff]

How important is it to have that much detail in the shadows?  If you look at the way you normally percieve a nightime scene, the lack of some detail in the shadows is not the issue that a lack of detail in brighter areas would be.   My experience with long exposures (30 sec or more) indicates that there is a surprising amount of shadow detail available if you use a levels adjustment layer and a mask to selectively bring up the levels in the shadows.

[feppe]

How important is it to have that much detail in the shadows?  If you look at the way you normally percieve a nightime scene, the lack of some detail in the shadows is not the issue that a lack of detail in brighter areas would be.   My experience with long exposures (30 sec or more) indicates that there is a surprising amount of shadow detail available if you use a levels adjustment layer and a mask to selectively bring up the levels in the shadows.
[a href=\"index.php?act=findpost&pid=85267\"][{POST_SNAPBACK}][/a]

Yeah, I do use masks and all that. I'm interested in squeezing out the most detail from the photo as possible as I can always lose it later - and usually do. Bringing back something that wasn't there in the first place is impossible.

Besides, the technique I described works as well with highlights to retain detail as I point out. Highlight detail is, indeed, a problem with long exposures as neon lights, street lights etc. tend to overexpose nearby areas.

[jani]

How important is it to have that much detail in the shadows?  If you look at the way you normally percieve a nightime scene, the lack of some detail in the shadows is not the issue that a lack of detail in brighter areas would be.

That depends on how "realistic" you want the scene represented.

Alain had a write-up on that particular problem just a few days ago.

In some cases, such as the samples posted by mahleu and Tim Gray in the Two exposures done in raw and..... thread in User Critiques, it is fairly easy to see how artistic vision can make use of those details that are "shadow" to the eye, but perhaps not to the camera set to long exposures.

[TimothyFarrar]

One way to think about stops of dynamic range is the following. Average sensor has 12 bits of precision from a mostly linear capture (excluding the upper highlights).

1. the brightest stop has 2048 levels (the highlights)
2. the next darker stop has 1024 levels
3. then 512 levels
4. then 256 levels
5. then 128 levels
6. then 64 levels (effective limit of usability in shadows)
7. then 32 levels
8. then 16 levels
9. then 8 levels
10. then 4 levels

So yes the digital camera captures more than 6 stops, but no they are not useable. Even if 16 levels was enough for good texture in the tonal range for that stop, it is much too corrupted by noise (even worse with long exposures).

Feppe as for your question,can you skip exposures and just grab a proper exposure (non-clipped for the highlights) and then another exposure for the darks to grab better dynamic range? Yes.

So say you captured 2 exposures 3 stops apart (going through the stops again).

1. 2048 levels in the darkest exposure, fully clipped in the lighter exposure
2. 1024 levels, fully clipped
3. 512 levels, fully clipped
4. 256 levels, 2048 levels in the lighter exposure
5. 128 levels, 1024 levels
6. 64 levels, 512 levels
7. 32 levels, 256 levels
8. 16 levels, 128 levels
9. 8 levels, 64 levels (effective threshold)

So if you can blend in such a way as to use the upper 3 stops from the darker exposure, then digitally drop the exposure of the lighter exposure 3 stops (8 times darker) and blend in the non-over-exposed regions of this exposure matched layer over the noisy shadows and mid-tones of the darker exposure, you will end up with the results you are looking for (less noise).Here is roughly how you could setup your layers,

 -> levels adjustment layer (set as a clipping mask on the lighter exposure)
lighter exposure [with layer mask to remove overexposed areas]
darker exposure [as background layer]

The key is making the levels adjustment layer a clipping mask (use OPTION+APPLE+G on the mac in CS2) so that you can use it to darken the lighter exposure to exposure match the non-overexposed regions to the darker layer below.

Here is an example of a slightly developed image (digital GND + contrast adjustments were applied as you will see below in the noise pattern),



Now lets look at the area between the white lines full size with a little sharpening and with bands with different amount of exposure blending (see below).

You will see that I threw in an undeveloped first exposure so that you could see how the digital GND was applied. Also my exposure 1 was exposed to the left to avoid any color distortion which is caused by non-linear responce at the upper end of the cameras sensor. Exposure 4 was not exposure matched correctly so blends that include exposure 4 are slightly lighter. Blending all exposures 1 through 6 (see column labeled "exposure 1-6") effectively removes all the noise. Using just exposures that are 3 stops apart ("exposure 1+4+7") is still a huge improvement over just a single exposure.



Or try http://www.farrarfocus.com/ffdd/blog20060926.htm for the full writeup on this test.

You can do the blending by hand. You will probably want to create yourself a Linear colorspace (Custom RGB... option in PS Color settings dialog) to work in so that the blending is correct, but you can still get good results using a non-linear space like ProPhoto RGB. Another key is to blend the images after loading them from ACR un-adjusted. Here are the steps to do the proper ACR settings, after loading both raw files at once in ACR and sychronizing the settings to apply to both images,

1. Adjust Chromatic Aberration, if necessary.
2. Adjust White Balance and Vignetting, if necessary. Even if White Balance is correct when set to Auto, you must adjust it up and then back down to remove the Auto designation.
3. Set Exposure, Shadows, Brightness, Contrast, and Saturation to zero.
4. Open the Curve Tab be sure that Tone Curve is set to Linear.
5. Set Space to ProPhoto RGB.
6. Set Depth to 16 Bits/Channel.

Try the blending process by hand, note you will still have to then develop the image afterwards in photoshop.

I've been doing the blending in an automated process (using a script in Photoshop) for a while now,
http://www.farrarfocus.com/ffdd/bracket.htm

[feppe]

Thanks Timothy for the very complete explanation! This query is actually based on the discussion we had on your FFDD email, and you continue to give priceless advice.

The levels example was a very good confirmation that I'm on the right track. I'm off to Paris in two days and will try this on the field. Then it's post-processing time (whenever I get back to my windows machine).

What's the reason for using linear color space? I work in ProPhoto RGB so isn't converting to another color space and back for blending purposes counter-productive, as smaller than ProPhoto RGB spaces lead to potential clipping?

[TimothyFarrar]

What's the reason for using linear color space? I work in ProPhoto RGB so isn't converting to another color space and back for blending purposes counter-productive, as smaller than ProPhoto RGB spaces lead to potential clipping?
[a href=\"index.php?act=findpost&pid=86431\"][{POST_SNAPBACK}][/a]

You are right in that it is probably better just to blend in ProPhotoRGB if you are planning on editing in ProPhotoRGB.