An eye tracker that doesn't need humans
This app uses a machine learning model to predict where users look. It presents the results as a heat map.
If you keep optimizing your digital product, at one point you may end up wanting to do an eye-tracking test. In 2021, most people are exposed to tens of thousands commercial messages every day, so every millisecond counts. But here's a problem: an eye-tracking test easily sets you back thousands of dollars, and if you want to own your own eye-tracker, it's not uncommon to shell out $20,000 or more.
The good news is that our eyes are not that unpredictable. What changes from study to study is what is being studied, not with what; human eyes. So the big question for a whole branch of AI researchers is, "can we teach a computer to look at an image as if it were a human?"
The researchers know that such a model would be useful for a wide variety of things. You could crop images without cutting out the main subject of the photos - like Twitter does in the feed. Cameras could auto-focus on what's important.
Today there are hundreds of these models. So how did they make it happen?
The mother study
A group of researchers at MIT invited 39 people, carefully selected to span 18 to 50 years old. They invited them on campus one at a time, and showed them 300 images, 3 seconds each. They were told they were doing a memory test, but they really were doing an eye tracking test.
It took 15 minutes per participant, that's it. Next step: make sense of the data. Come up with an algorithm, a machine learning model that could take a new and unseen image and predict eye movements.
Researchers all over the world are chipping in, comparing their results on a global leaderboard, with lots of them at an accuracy of more than 80%.
A precision in the high eighties might not sound perfect, but consider the alternative: Buy a $20,000 machine, get it shipped, learn how to use it, recruit 39 people, book a room, find someone who can take their calls, cancelations, brew coffee, offer coffee, brief them on the study. Repeat most steps 39 times.
And then there's the opportunities it unlocks that could never have been done with a traditional study, like cropping millions of photos, day and night, in less than a second per image.
Millions of years of data
There's a reason the MIT researchers only invited 39 participants. It's widely known as the sweet spot for eye tracking studies. If you invite 80 or 800 more people, the results are just not going to get much better. We're simply not that different when it comes to how our eyes are wired. We like contrasts, faces, and things that look weird. In other words, hand-me-down experience from thousands of generations of humans before us. The stuff that helped us stay alive when we were facing dangers like elephants and thunder.
The data this kind of study collects is what researches call bottom-up. It's almost purely biological, and almost completely free of cultural influences. A regular eye-tracking test would produce a mix of the two, and this is an important part of the explanation of the 80% accuracy.
And then, after millions of years of human eye development, and a decade of the world's best researchers studying the movements of said eyes, we could make this app. You can drag in an image, and even connect your iPhone and see eye tracking results in real time.
The app is currently in private beta, but you can get a copy if you email morten at otato dot app.