Image Distribution Surprises (The MIT 5k dataset 8)

by Dan Margulis on December 7, 2018

The MIT study offers a unique opportunity to study the distribution of images. Most of the images that I and other authors collect for our own purposes test or prove a certain point. The MIT dataset, on the other hand, simply gathers together images that might be worked on professionally, whether dull or interesting, without respect to topic or category. For my own testing I chose 150 of these images at random. That’s enough to draw valid conclusions as to how often we see certain things.

Granted that these 150 images likely represent something like the distribution one would find in a professional retoucher’s real life, I can’t say that I was too shocked by the distribution of subjects, which I listed here, nor was I particularly surprised by the finding in the same post that PPW gets superior results close to, but not quite half the time. The MMM action was valuable in well over half of all images. On the other hand, it made a decisive difference only in around half of that half. This led me to ask myself why it was so effective in these images and less so in others. I learned a lot from the answers and in the next round of documentation I hope add this information.

Let’s start with a quiz based on these selections.

*How often do you think an image divides neatly into two parts, one more important than the other? I will call these “foreground and background” because that’s what they usually are, but any other obvious division into two parts is acceptable. Furthermore, for the sake of argument let us say that whichever half of the image is clearly more important shall be referred to as the foreground, even if it is in fact the background.

*In images like these, professionals often enhance the foreground in a way that costs detail in the background. This is considered sound practice, guiding the viewer’s eyes to the more important area. But how often is this undesirable? In other words, how often is the background interesting enough in itself that we should be reluctant to damage it?

The answers, and why they are important, will come at the bottom of the post.

You don’t have to work all that many years in professional retouching before you start to think you’ve seen everything. And, indeed, as far as image challenges goes, it’s hard to imagine now that I could face something completely surprising. Certainly, no single image in the MIT study presented a problem that I hadn’t seen before many times.

There were several cases where the retouching team did unexpectedly badly and a few where I simply screwed up. These cases prove only that we are human beings and have nothing to do with the merits of PPW.

A couple of results surprised me, where I was expecting an easy win for PPW and got something less agreeable, learning something from the experience. I’ll show those in a subsequent post. This time, I’ll talk briefly about the how often certain problems come up. This is an underappreciated topic. When we discover a method that works exceptionally well, it’s easy to overestimate its importance. Who knows how often images come along that would benefit from it?

The obvious example is the Bigger Hammer action. This is our most powerful means of enhancing highlight detail. At its best, it is staggeringly effective. At conferences I work with an image of Niagara Falls and draw gasps from the audience when they realize how it achieves what seems impossible.

That is the sort of this that lulls me into believing something is more important than is actually the case. Of these 150 images in only two cases do I believe that the Bigger Hammer made a serious difference. One of them, the fountains of the Bellagio, is shown in this post. The other was a night scene where the action got noticeably better detail in areas under street lighting. So I’m glad that the action was useful but it hardly is the kind of thing to create gasps of astonishment. I’m happy to have the Bigger Hammer available just in case a waterfall or tsunami image comes long but this study made me realize it is not one of the most important actions in the panel.

Meanwhile, two milder actions, the false profile/multiply set and the Velvet Hammer, proved their merit on many more occasions.

Back to the quiz. For around thirty years I’ve been telling people that they make too many selections. The argument has gotten very complex, with a lot to say on both sides. The origin of my statement, however, was sound. Back in the prehistoric days when computers were very slow and Photoshop not very capable, retouchers still had to work on images that involve a foreground object that the client is greatgy interested in, and a background of lesser importance. My teachings back then were: apply curves that accentuate the foreground. If this costs detail in the background so be it; this is how humans would visualize the scene anyway.

The counterargument: now that computing time is no longer a factor, why damage anything? Why not deselect the background, so that the foreground improves but the background doesn’t change for the worse? I have always replied that the change for the worse is actually an improvement that makes the image more natural and guides the eye to the proper place, and that the client generally prefers moves that accentuate the foreground.

The question was revived in recent years. The MMM action works as described above: a certain area is designated as representing the important part of the image, which is then enhanced at the expense of other things. Gerald Bakker, a student of PPW, came up with a counter-action he called MMM Finetuned. This action, which we incorporated into PPW panel v5, allows enhancement of multiple areas. It has many applications but the most important, in my opinion, is when the image does not match my description above because there is an objection to losing detail in the background. This defies the theoretical principle that concentrating on one area logically implies less concentration on another. Instead, it suggests that there is something so interesting in the background that the viewer should be given the option of appreciating it. This is the principle behind “High Dynamic Range” imaging: instead of improving the foreground at the expense of the background, improve both at the expense of neither.

The quiz, then: of 150 random images, how many neatly divide into two pieces of which one appears more significant than the other? And of these (call them foreground-background shots, even though it could be the background that’s more important) how often is the less significant piece nevertheless interesting enough to justify trying to hold or even augment its detail?

My guess at the outset was that two-thirds of the images—100, that is—would be classified as foreground-background, and of these perhaps 10 or 15 would have an interesting background. This was slightly too conservative: the final count was 102 and 21. So, Gerald’s action has more utility than I would have predicted.

{ 0 comments… add one now }

Leave a Comment

Previous post: