Google has rolled out a new feature in its Allo messaging app that converts your selfies to cartoon-like personal stickers.
It does this by combining its neural networks with the work of artists to automatically generate customizable sticker version of your selfies.
Rather using computer vision to analyze individual pixels, Google searched for an algorithm that was good at picking out features of a person's face, allowing it to bypass constraints caused by lighting conditions. Rather than build a new neural network, its experiments with existing general purpose computer vision neural networks uncovered certain neurons among these networks that were good at focusing on facial features, despite never having been trained to do that.
It relied on artists to provide a set of features such as hairstyles for customization. Then, with the help of human raters, it used the illustrations to teach the network to match the different hairstyles with selfies. The human raters also judged the quality of the output compared with the selfie.
Google also wanted to avoid the problem the "uncanny valley", the sense repulsion people feel when looking at a copy of a human that looks almost alike, but not exactly the same. There are plenty of examples of it in realistic animation.
"Rather than aim to replicate a person's appearance exactly, pursuing a lower resolution model, like emojis and stickers, allows the team to explore expressive representation by returning an image that is less about reproducing reality and more about breaking the rules of representation," explains Google Allo's expressions creative director, Jennifer Daniel.
Customization options include hairstyles, skin tones, and nose shapes. Google has attempted cater to a range of identities, offering users a way to express race, age, masculinity, femininity and androgyny.