Thursday, 17 December 2015

This Algorithm Can Tell How Memorable a Photo Is

memorabilityhead

Some photographs have the power to burn themselves into our memories for a long time, while others are easily forgettable after they’re seen. Scientists are MIT have been researching the science behind memorable images, and now they’ve created a web app called LaMem that can analyze any photo and assign it a memorability score.

lamemscreen

The demo was created by the folks over at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), which created a deep learning-based algorithm called MemNet. They say that MemNet is able to predict the memorability of photos roughly as well as the human brain.

Scientists began the project by doing a crowdsourced experiment with 5,000 online participants who were each shown a set of photos and asked to press a key whenever they saw a familiar image. Scientists took the data and used it to create a memorability score for each of the images.

research

Although the participants were diverse in their backgrounds, they generally agreed on which photos were memorable and which weren’t.

results

The known photos were then used to train an algorithm to make its own predictions, which were found to be 30% more accurate than existing algorithms. LaMem even generates a heatmap of each photo to highlight the areas that are considered memorable.

Here are a few results from example photos from sample photos:

memorable1

memorable2

We did a couple of iconic photos as well (by photographers Dorothea Lange and Ansel Adams):

memorable3

memorable4

In addition to helping photographers create more memorable photos, marketers can also use the tool to optimize the memorability of their advertisements. You can learn more about LaMem and read the research paper for yourself here. Head on over to the demo if you’d like to find out the memorability of your own photos.

(via LaMem via Engadget)



from PetaPixel http://petapixel.com/2015/12/17/this-algorithm-can-tell-how-memorable-a-photo-is/

No comments:

Post a Comment