These days, artificial intelligences can modify, transform, and even invent portraits. But a group of researchers from Adobe Research, UC Berkeley, and KU Leuven propose something different: erasing our arms. The Unselfie project aims to adapt traditional selfies so that the final image presents a more neutral and natural pose. An "algorithmic selfie stick", if you will.

Unselfie: An Artificial Intelligence That 'Erases' Arms from Selfies
Artificial Intelligence

For better or worse, the selfie has already become a fundamental pillar of modern photography. From basic smartphones to cameras worth thousands of euros, almost anything with a lens has the potential to take a selfie, no matter how strange it looks (and I say that remembering people who posed using 12-inch tablets).

Over time, users began experimenting with selfie sticks and other accessories to improve their self-portraits; however, there is something not entirely natural about the final result, and a group of researchers from Adobe Research, UC Berkeley, and KU Leuven (Belgium) believe they can correct it with pure artificial intelligence. The project is called 'Unselfie', and in simple terms... it erases our arms.

How Unselfie Works

From a technical standpoint, the Unselfie algorithm was trained with more than 23,000 'neutral' photographs and more than 4,600 selfies, with the goal of 'relaxing' arms and shoulders, so that the selfie conveys the feeling of having been captured by a third party, or with the help of a timer.

Unselfie: An Artificial Intelligence That 'Erases' Arms from Selfies
The results of Unselfie, on the right. Not bad at all.
https://old.neoteo.com/pinturas-famosas-con-rostros-reales-y-animados-como-se-verian/

This transformation also extends to the torso, neck, clothing details, and of course, the background. "Filling the holes" is one of the most complicated parts in the training of the artificial intelligence, along with the lack of pairs (that is, a photo as a selfie and another neutral one taken by the same person).

After several failed attempts, Unselfie's proposal is to apply a pair synthesis system with self-supervised learning. The results are far superior compared to other similar methods, especially in terms of detail preservation and landscape adaptation. Would you like to try it? Well, the truth is you can't... but it's only a matter of time before someone turns Unselfie into a web service.

Access the study (PDF): Click here