An AI can make selfies look like they’re not selfies


There’s a solution for tourists who are reluctant to hand over their expensive camera phone to a random passer-by to snap a photo. A new computer model promises to make a selfie look like it isn’t one.

The technique borrows from the growing field of artificial intelligence that can modify, or ‘repose’ photos in a realistic way, says Liqian Ma of KU Leuven, Belgium. Reposing uses two or more source images – one in the original position and another in the reposed position – to train its algorithms. But selfies only have a single source image: if it was possible to repose the subject to stand in a portrait taken by a third party, there would be no need for the selfie.

The “unselfie” computer algorithm trawls through a database of portraits, taken from collections of fashion stock images, and uses the information to create a computer-generated model of the person in the selfie. It then searches through the database of portraits to find one with a pose most similar to the pose desired in the modified selfie, before using the computer-generated model to map the original portrait onto the new pose.

The algorithm then trains itself to get better and better, and fill in the pixelated gaps that occur. “It’s very sensitive to the imperfections of the generated models,” says Ma. Interviewees gathered through Amazon’s Mechanical Turk system judged the resulting “unselfied” images better than reposed images created using other AIs developed previously.

The paper uses a “clever approach” to solve the problem, says René Schulte, an AI and spatial computing developer for German company Valorem Reply. “It still has limitations and can still result in unpleasant artefacts in results, but it surely is better than previous approaches.”

Ma plans to train the model further to tackle those problems, incorporating a broader range of poses and matching the background to the new limits of the reposed human more smoothly., 7 August 2020
; ”>