Researchers from the University of Texas at Austin, Cornell Tech, Georgia Tech and Facebook AI Research have developed an artificial intelligence (AI) app that can give fashion advice. The AI system can look at a photo of an outfit and provide helpful tips to make the outfit more fashionable.
(Source: Kimberly Hsiao/University of Texas at Austin)
The team was motivated by the idea that a person can take a given outfit and make it more fashionable with small adjustments. People can create a fashionable outfit without purchasing any new items.
The app is named Fashion++, and it uses visual recognition systems to analyze color, pattern, texture and shape of garments in a given image. The AI system considers where outfit edits will have the most positive impact on the outfit. After analyzing the image, the system can offer tips and several alternative outfits for a more fashionable look.
Fashion++ was trained with more than 1,000 images of outfits from public online websites for fashion enthusiasts. Because these sites are well-curated, the team struggled to find examples of bad fashion to train the AI system. To overcome this issue, the team mixed images from the website with other fashionable images to create an unfashionable look. The team then trained the system on what not to wear using these images.
The AI can continue to learn what is fashionable over time by giving it updated images of fashionable clothing, which is abundant on the internet.
The team admits there was some bias while creating the system. Vintage looks were harder for the system to register as fashionable. This is because the images used to train the AI were from the internet, and the internet has only been in wide use since the 1990s. Most users who submitted to these sites were from North America, so styles from around the world are less likely to be suggested as fashion options. Often body shape and size are major factors in fashion choices. But many of the fashionable images the system studied were worn on a model’s body, which limits the system’s ability to give fashion advice based on body shape and size.
The team is working to overcome these challenges. They plan to introduce a wider variety of body shapes, sizes and international fashion trends to diversify the system. They want the system to offer tailored recommendations for individual users, no matter their location or body shape.
The paper on the system can be found here.
