Large-scale Image Memorability: How memorable is your image? Upload to find out!

http://memorability.cs​ail.mit.edu/demo.html This is a Convolutional Neural Network trained on image memorability studies. Just going thru some random images I've made, I found a couple that scored very high on the memorability scale.

First example: a random macro photo.

The second example: a CGI logo.

The color overlay represents the computed normalized memorability for regions of the image, in case you wondered.

So... how do your images fare?

What gives an image this undefined memorable quality wasn't clear, but from your examples I thought the whatisit was going for large areas of color contrast. I therefore chose these jar lids.

OhLook wrote in post #17825316
...

Hosted photo: posted by OhLook in
./showthread.php?p=178​25316&i=i39928549
forum: General Photography Talk

...

Score!

Trying for a low score this time. This image is pretty homogeneous. I predicted it'd be forgettable.

Let's see, here's a rather ordinary shot of a chicken.

Well, with a score of 100% memorability, that gives a nice control for experimentation.
First I tried removing some of the little flowers:

Then I tried an indigo look:

Then I added a heavy USM:

The larger the subject is, and the more it's separated from the background, the higher the score. (among other factors, I guess like color!)

It's really a failed app, as some pretty lame images I have sent up got "very high" where as the images that I have sold the most of in gallery shows, got "high" or less. Not sure what the thinking is that gives them the idea to use the term "memorable"

It really scoffs at landscapes:

CyberDyneSystems wrote in post #17825347
The larger the subject is, and the more it's separated from the background, the higher the score. (among other factors, I guess like color!)

It's really a failed app, as some pretty lame images I have sent up got "very high" where as the images that I have sold the most of in gallery shows, got "high" or less. Not sure what the thinking is that gives them the idea to use the term "memorable"

From the paper: "...memorable and forgettable images have different intrinsic visual features..."* This app produces result that are on par with human results (64% vs 68%). It's not the margin of error that is fascinating about the app, but how close to human error they got.

*Understanding and Predicting Image Memorability at a Large Scale
A. Khosla, A. S. Raju, A. Torralba and A. Oliva
International Conference on Computer Vision (ICCV), 2015

Trying a B&W shot with much small detail:

I tried a couple of my images from the recent thread, WORLDWIDE PHOTO WEEK: Part 468 - Minimalist

First, an 8x8 square of black pixels...