Picking out clothes that work together and look fashionable can be a difficult task for many of us. We may have our own sense of style but a sense of fashion is different altogether and it can be tricky to choose clothing combinations for particular events or occasions. But now a team of computer scientists have come up with an algorithm that can aid those in need of fashion help.
Scientists Raquel Urtasun and Sanja Fidler from University of Toronto, along with colleagues in Spain, have created an algorithm which can look at photos of an outfit and suggest ways to make it better. They presented their paper detailing the software at the Computer Vision and Pattern Recognition conference in Boston, Massachusetts. The machine learning algorithm works by visually parsing the image and evaluating the clothes themselves, but also by analysing the appearance of the person wearing the clothes, the ‘type of outfit’, the person’s surroundings and metadata like the country and city of the person in the photo.
The algorithm uses the Orbeus ReKognition API to scan facial features and attributes, then uses information from Flickr80K to detect different image styles. The software goes on to use information to decide on the ‘fashionability’ of the outfit, based on previous input from public users about what they have believed to look fashionable in similar settings. The algorithm creates a complex equation which it uses to suggest ways in which the outfit can be improved. The paper says that the ‘Conditional Random Field’ model that jointly uses a number of factors to create results “significantly outperforms a variety of intelligent baselines”.
The researchers used the community of chictopia.com to make recommendations on what they thought was fashionable or not fashionable. It has been acknowledged that fashion preferences can vary based on a number of things such as taste, nationality and gender, each of which can play an important role when analysing images. However, Chictopia allowed the scientists to gage opinions on what is generally considered fashionable and what isn’t, not taking into account the various personal tastes of the users. The site encourages global users to post photos of their daily or nightly outfits along with information about the clothes they’re wearing and where they were bought or acquired. Others can then vote for the outfit to show that they like it and think it’s fashionable, as well as being able to leave comments about the photo. The scientific team collected 144,169 posts from the website and analysed them in order to find out the specific characteristics that encouraged approval from users and the characteristics that didn’t.
The authors wrote in the paper “Our aim here is to give a rich feedback to the user: not only whether the photograph is appealing or not, but also to make suggestions of what clothing or even scenery the user could change in order to improve her/his look”. The data that was collected by the scientific team also enabled them to rank cities in order of the most fashionable users- Manila topped the list while Jakarta found itself at the bottom.
The algorithm is capable of advising people on how best to improve their outfits but also benefits people interested in taking better selfies and OOTD (outfit of the day) pictures. Chictopia could be seen as providing a limited set of information that may not generalise to the whole population, however, because the site members are made up almost exclusively of young women with a small number of young men alongside them. The scientific team is hoping to use other sources to capture data in order to further refine the results of the algorithm.
There are not yet any reports on what the algorithm will be used for in the future, but we may see it emerge as part of a smartphone app that can be used easily by individuals when they enter images of their outfits. As well as being able to help those who have trouble with their fashion sense, the algorithm could have commercial applications which will make it useful. Trend analysis is an area where such software could play a part and would likely be used by large fashion brands and forecasting companies.