The characters drawn by children can jump, walk and dance, just like they do in the animated series. Meta researchers have created a tool to animate hand-drawn graphics, as long as they feature a person’s characteristics, such as two arms, two legs, or a head. In a matter of seconds, the program turns a drawing on paper into a character that moves up to 35 different ways.
It works like this. First you have to take a picture of the original drawing and upload it to the platform page. The program detects the drawn figure and separates it from the environment, highlighting the character. At this step, adjustments can be made manually (using the stylus to add and the rubber to erase). After isolating the figure without including anything else on the page, a simplified structure of the body appears above the drawing, indicating what the head, torso, arms, and legs will be. Those are the joints of the personality, so that later it moves as if it were a person. And if the character doesn’t have arms, for example, that’s no problem: the program suggests pulling “the elbow and wrist joints away from the character” to make it move as if it were. Once you’ve completed all of these steps, your drawing will be ready to jump, walk, wiggle, or dance. At the end of the process the animated video can be downloaded and shared with friends, family or on networks through a link.
The creators of the tool argue that the vast majority of artificial intelligence (AI) technologies are used to analyze realistic images of people, such as photographs. But the way children draw, often in an abstract, asymmetric, and unrealistic way, makes this identification difficult for the machine. Jesse Smith, a research engineer at Meta, explains to EL PAÍS that there is a lot of variation in the style and morphology of this type of figure, which makes it difficult for AI to make accurate predictions. “It is also difficult because many of the keys useful for analyzing objects in pictures, such as colors and textures, are not found in amateur graphics,” he asserts.
While a parent or teacher can easily identify a bee with legs and shoes, a tree with an arm or a chariot with a face that represents a human figure in a child’s imagination, it is difficult for an AI to find this assimilation. Doing so, the engineer explains, requires a lot of data. “There were no large datasets to train the algorithm on. That’s what prompted us to launch this project,” says Smith.
The form was developed over a year ago and 6.7 million images have been uploaded to the platform’s beta. Last week, Mark Zuckerberg, founder of Meta, announced the opening of the code and database of 180,000 drawings for researchers and other creators to use. “We published a peer-reviewed technical document that describes in detail how everything works and all the design decisions,” says the company’s engineer.
While the show is mostly entertaining, it can also be useful in the educational field, such as talking about technical terms from an early age, in a way that children can easily understand. “I know teachers who use a browser demo to explain machine learning: A drawing is an input. The model thinks the elbow is here and that’s a prediction. See how it’s a bit out of place? That’s a prediction error,” he explains via email.
Is it possible to add other features to the animation, such as sounds, to have a video story that combines drawing and dialogues? Smith confirms that yes, though Meta has “no plans” to implement them at this time. “However, since the code and data are open, others could create something like this,” the researcher continues. “I have a 14-month-old son. By the time I turn 5, I’d like this to be part animation storytelling tool, fully featured for easily creating your own animated stories from scratch.”
You can follow The Country Technology in Facebook And Twitter Or sign up here to receive The weekly newsletter.