Articles

How Memorable Are Your Images? MIT Algorithm Predicts How Memorable — or Forgettable — They Are

We spend a lot of time — or not — composing our images, taking the photos, then adjusting them in post-processing to create the kind of images that we hope will wow our audience. Who doesn’t want their images to be memorable?

Now, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) believe their “MemNet” algorithm can predict how memorable — or forgettable — your images are. They even plan to turn it into an app that will subtly tweak your photos to make them more memorable. Everyone, it seems, may soon become a great photographer.

These researchers study memorability — how likely it is that someone will remember a visual message. Besides improving your photos, other practical potential applications may include improving the content of ads and social media posts, developing more effective teaching resources, designing better logos and products, and even building personal health-assistant devices to help you remember things that are most important to you.

To aid in that research, the team has amassed the world’s largest image-memorability dataset, LaMem, comprising 60,000 images, each annotated with detailed metadata about qualities such as popularity and emotional impact. You can try MemNet online by uploading your own photos: it will create a heat map that identifies exactly which parts of your image are most memorable. It also gives you back a score ranging from 0 to 1, where 1 is the most memorable.

mit-lamem-memnet

I ran a photo of tulips I took through the algorithm and it accurately singled out the tulips from the foliage around it. It returned a score of 0.577, which means that it expects that about 58% of viewers will still remember that image 100 seconds after viewing it. Browsing the LaMem dataset, it is perhaps not too surprising to find that images with people, vibrant colors and interesting shapes in them are rated the most memorable while landscapes and real-estate type images rate as the most forgettable.

I guess if this algorithm were to be included in a camera and it gave live feedback on your composition, it could alert you to the fact that some parts of your photo may be competing with your main subject. It will be interesting to see if MemNet makes it into an app and the app makes it into your camera.

Try MemNet online