D800E vs. 5D3: Diminishing Returns or Reversal of Returns?

Started Jul 19, 2013 | Discussions thread
The_Suede Contributing Member • Posts: 644
Personal opinions, Scaling&Sampling theory continued

To start with, I'd like to observe that it's often very important to use the few simple words "in my opinion" or "as I see it" when you're disucssing "soft" issues, i.e issues that for most non-image-scientists cannot have a clear, uncut and scientifically based answer.

It just saves a lot of time, since you then disarm the "you are WRONG!" answer. It changes that answer into "I don't see it that way" - which is a much less emotionally tense situation.


Regarding the sampling theorems, I see in practice what the sampling theorems say "should" happen.

Detail DOES get lost when downsampling an image by a large factor. There are occasions and scenarios where this isn't true, but for normal imagery it generally is. There are two main reasons for this.

1) The sampling algorithms we use in PS are fairly primitive. When you scale an image down below the "pixel perfect" size (*1) by a factor three or more, significant detail loss occurs. The convolution kernel used in PS (to get speed) is heavily biased towards the center of the sampling window, meaning that when you pull the sampling points apart (downsample by >3x) the information in between the new pixel centers get more and more ignored. You can see this as aliasing and moire, and it gets worse the more you downsample per step. Going directly from 36MP to 600x400px can make the image look very strange.

2) Your image is both S-curve and Gamma converted, as any normal photography should be. But from a resampling PoV, this is a disaster... Lets say you have two grays, one dark (RGB=20) and one bright (RGB=220). When you mix them, the result will be "120". But that's totally wrong, the real answer (when mixing sRGB gamma images) should be "162"... The 120 result will be to dark due to the gamma curve, and this will affect how details render.

(*1) No Bayer-based digital camera will be "pixel perfect" at 100% scale. You can get close, depending on the amount of color /chroma) detail in the target you shot, but never 100%. Most normal applications get to "very good per-pixel detail" when downsampled to about 70% size.

So depending on how well the Bicubic kernel plays with a 22MP > 1800x1200px conversion vs doing the same from 36MP, that may change the outcome significantly (significant, here: Most people will see it if looking carefully)


Ill just show two images that shows a bit of what Bicubic and sRGB gamma does to resampling. The first one is just my example of blending 20 with 220, and getting 120. That sRGB result is to dark, it should be like on the left side where I've done the average in linear space in stead.

mixing or averaging gamma treated values

This error might not seem very important, but when you combine this with the quite erroneous results you get from Bicubic sampling on small detail, results get strange. As you can see, the lower part (right) where I've taken the "original" and downsampled to 50% with normal PS bicubic is much to dark in the center (where it should have been medium gray), and this also amplifies aliasing effects. The upper part (right) was also downsampled to 50% in PS, but on a linear gamma original.

This is the result we want. The smalle version should look identical to the larger version, only SMALLER - and it will contain less detail per image-width of course.

The image must be viewed at original 100% size! The downsampling algorithm DPR uses mess up the result even more than PS....

Post (hide subjects) Posted by
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow