To teach the software about fashion, Urtasun's team showed it thousands of pictures from Chictopia, a popular style website. The more positive votes left by other users, the more "fashionable" the software perceived the look to be.
It also noted other information about the photo, such as the user's geographic location, the date they had posted it, the background of the picture, and written descriptions of the clothing.
The resulting software uses this information to categorise outfits and make suggestions based on what was successful for others in similar situations – for example, to add black boots or try something in pastel. The team plans to hone the results further by showing it a more diverse array of photos from other sources.
Urtasun presented the work at the Computer Vision and Pattern Recognition conference in Boston, Massachusetts, earlier this month. Her team plans to improve the software so that it can automate the work of a human stylist.
Alexandra Greenawalt, a personal stylist in New York City, is understandably sceptical about computers muscling in on her patch. Looking good is about more than the latest trends, she says.
When dressing clients, she considers a wide range of factors, including their age, occupation and body shape. An effective algorithm would need to take all that into account, too.
Still, she is curious to watch the technology develop. "What will be interesting to see is if it can predict fashion before it happens or just based on likes in the past," she says. "I would imagine the teens and 20-year-olds who are very much wanting to be in fashion would find it valuable."
No comments:
Post a Comment