Technical Perspectives on Digital Photography

There is a bit of a debate in some of the photography forums on whether the ISO 3200 setting (enabled through a custom function) on some of the Canon bodies is a true ISO 3200 or if its simply the camera underexposing by a stop at ISO 1600 and then digitally pushing a stop to get ISO 3200. It has also been suggested that RAW converters like ACR and Lightroom can do a better job of this digital pushing than the camera can. I was curious to find out if this was true and so did a few experiments. I will walk through some of the data I collected and how I collected them, so if you are interested only in the conclusion then skip to the end of this post.

First let me say that these results apply only to the camera I tested with (the Canon 5D) and only to Adobe CameraRaw; results may vary with different cameras and/or different RAW converters. I used two test images for my experiments. The first was a Macbeth color chart (so that I could easily notice any color shifts) and the other was my favourite fake plant. For all the RAW conversions I set the 'sharpening' and 'color noise reduction' sliders to 0. All tests were done in 16-bit in ProPhotoRGB.

I took the mean of 5 images at ISO 100 to compute the base-line against which the shots at the other ISOs would be compared. I took the difference between the base-line and the image for each ISO and then computed the mean difference and standard deviation over the entire image. The mean difference tells us on average 'how close' an image at a particular ISO is to the base-line and the standard deviation gives us an idea of how far spread out the pixel value differences are in reference to the base-line image. Here is the result for the Macbeth color checker:

The results are rather curious. It seems that for ISO 1600 digitally pushed the mean difference is lower where as the standard deviation is higher. I then proceeded to plot the same graph for the 'lily' test scene.

The overall trend is confirmed in the 'lily' test scene as well. Also interesting to note is that the mean differences for green is lower than red which makes sense due to the Bayer pattern and is something that is quite desirable (as the human eye is more senstitive to green than blue in well lit conditions).

So how can we interpret this data? Since the mean pixel difference is lower for the digitally pushed ISO 1600 image, I would say it is a truer representation of the actual scene. However since the standard deviation of the pixel differences is higher for the digital pushed image, it means that there are parts of the scene where the differences are larger. It would be interesting to learn which parts and to do this I did a qualitative examination of the images.

From the small crop of the Macbeth color chart image we can see that the pushed ISO 1600 image has more differences in the dark or shadow regions where as the ISO 3200 image has a lot of difference in the highlight area and not as much in the shadows. Here are the curves adjusted differences for the 'lily' test image:

ISO 1600 digital pushed


ISO 3200

Again in our 'lily' test scene we notice that ISO 3200 performs better in dark/shadow areas and ISO 1600 digitally pushed performs better in highlight areas. Since these two scenes had a lot of highlight regions, it explains the statistical results from above. One way to confirm this hypothesis would be to make the same comparison on a test scene with a lot of shadow data and that is exactly what I did next.

We can now see that in this scene both the mean pixel difference values and the standard deviations are considerably higher for the digitally pushed ISO 1600 image vs. the ISO 3200 image. We can definitely conclude that for shadow areas ISO 3200 produces a better result.

There was one last thing I was curious about. I began to wonder if this was something unique to ISO 3200. I tested this by comparing the mean and standard deviations of differences between an image at an ISO value vs. one of one stop ISO less pushed digitally. The following two graphs are the results:

As you can see the differences are large at all ISOs. The difference diminishes slightly at higher ISOs (before picking up again at ISO 3200). I then plotted the mean pixel difference for each pushed image vs. native ISO 100.

Again we see that the mean difference is high and is slowly (very slowly) decreasing. This to me indicates that there isn't anything THAT out of the ordinary going on at ISO 3200.

A lot of the differences in the values has to do with the fact that Camera Raw tries to save as much shadow information as possible during its conversion. This can sometimes manifest itself as shadow noise, but you need only to move up the black point to get rid of this noise.

So in conclusion, what should we do, ISO 3200 or push digitally. I'd say if having less shadow noise is important then shoot at ISO 3200 but if the scene is well lit and you don't have too many dark areas then shooting at ISO 1600 and pushing digitally might be the way to go. After all you can always get rid of that shadow noise (either by moving up the blacks or with noise reduction).

Several months ago, I was in the process of adding better color management for an open source rendering and simulation package. One of the things that many schools don't teach in their rendering courses is proper color management (along with proper gamma correction). Most people when writing their first ray tracer (in school or elsewhere) usually represent their colors as RGB without giving too much thought to which specific RGB color space (or even if the color space they are using is linear or non-linear). Generally, you want to do your intermediate rendering color computations in a wide gamut linear color space (I suppose you could do it in a different colorspace such as CIE XYZ or CIE LAB, but I wouldn't recommend doing things like BRDF calculations in such color spaces (however if you are doing spectral integration, you want to integrate to CIE XYZ, then convert to your internal RGB color space)) in high precision (preferably floating point for higher dynamic range). Naturally these things are a tradeoff against space and speed. Personally I'd rather be correct and accurate rather than fast, but YMMV.

So upon deciding that I should change the internal rendering color space to ProPhoto RGB I set off to find out as much as I could about ProPhoto RGB. After several online searches I came up with a few descriptions of the color space such as the one by Michael Reichmann, the one on Nature Photographers online magazine and its Wikipedia entry. Sadly, after hours of searching I couldn't find any technical details on the color space such as its white point, chromaticities or its nonlinear transfer function. I did manage to discover that it was first introduced and proposed by Eastman Kodak, and so went off and did another round of searches with 'ProPhoto RGB' and 'Kodak'. Again, not much luck. I finally gave up and asked a colleague what gives. It turns out ProPhoto RGB is the same as ROMM RGB. ROMM RGB was developed by Eastman Kodak, however I believe the name ProPhotoRGB is used because of the friendliness of the name (I guess ROMM RGB just doesn't have the same "pro" ring to it).

Now searching for ROMM RGB, led to easy and quick success. I was able to easily find the original white paper along with the specifications I needed.

Also, if you open up Photoshop you will notice that you can assign and convert to both 'ProPhoto RGB' and 'ROMM-RGB'. Though they are named differently, they are in essence the same.

One of the things shadow/highlight adjustments are useful for is to correct for underexposure in a part of a picture. However this feature seems to get overused (and hence abused) a lot; which results in (what I consider) to be nasty halo artefacts.

Here is an exampe for the kind of picture you may want to jazz up using shadow/highlight:

Now if you run shadow/highlight on this image and realizing you want your sky to be a little darker, get a little too liberal with the highlight processing, you end up with something like this:

The halo-ing at the horizon to me tooks displeasing. However these kinds of images seem to pop up everywhere; especially with people playing around with HDR. Now I wouldn't presume to dictate what is and isn't art, hence I will concede that this type of shadow/highlight 'abuse' could be used to render an image with specific artistic intent.

If you did want to brighten up the shadow region in this image, by being a bit more judicious in the application of highlight processing, we can avoid some of the halo-ing artefact:

Incidentally, the correct answer in such a situation would be to take series of exposures and merge to HDR. Once you have an HDR image, you'll want to convert back to LDR using the appropriate exposure and gamma values that gets you the desired detail in both shadow and highlight regions. However be wary as some HDR->LDR tonemapping algorithms can introduce similar artefacting.