Epic Challenge of "What to Wear Today?" Easily Solved by Artificial Intelligence...

2019-10-28

It's that time of year again, when the seasons change from autumn to winter. Every morning, after struggling to get out of bed with all my determination, I am immediately struck by the question of "what should I wear today?" After half an hour of indecision, trying to find inspiration for my outfit, I finally put together an ensemble, only to feel that something is not quite right... For many people, fashion coordination is a difficult subject to master. Besides finding a style that suits oneself, one must also understand the characteristics of each individual item in order to create a visually appealing outfit. How many girls have dreamed of having a personal stylist who can effortlessly put together their outfits every morning, relieving them of the torment of the century-old question of "what should I wear today"? Well, your wish has been granted with the development of an app called "Fashion++" by the Facebook team. This app helps you bid farewell to excessive shopping (of course, I know this is impossible) and fully utilize your existing wardrobe to instantly transform into a fashionista. In the movie "Clueless," the fashionable protagonist Cher Horowitz had a virtual stylist who allowed her to preview and select her favorite outfits. Now, this futuristic scene from the movie has become a reality with AI technology, and it has even surpassed our expectations. Facebook's latest personalized fashion coordination recommendation model, Fashion++, cleverly adjusts existing clothing items through algorithms to make them more fashionable. Even fashion novices can instantly transform into fashion experts. Fashion++ is a neural network that uses a deep image generation model to recognize clothing and provide suggestions for modifications. For example, the model may suggest removing, adding, or repositioning certain parts of the clothing. It can also modify specific aspects of the clothing, such as tucking in a shirt or rolling up sleeves. Unlike previous studies that directly propose modifications to the entire outfit or provide similar alternatives, Fashion++ modifies existing clothing items to make them more stylish. AI provides fashion modification suggestions in the following manner: In the left image, AI suggests removing the shawl from the overall outfit. In the middle image, AI suggests replacing the women's shirt with a sleeveless one. In the right image, AI suggests rolling up the sleeves of the shirt. Fashion++ does not encourage excessive spending. It maximizes the use of your existing wardrobe and provides specific suggestions for fashion coordination. This model consists of a deep image generation neural network that recognizes and learns to synthesize clothing based on the encoding of each item. The encoding explicitly decomposes the shape and texture of the clothing, allowing for direct editing of fit and color/pattern/material. The overall architecture of the model is shown in the figure. Researchers first obtain implicit features from the material and shape encoders. Then, they use the editing module F++ for calculations. After one round of editing, the shape generator decodes the updated shape features, restoring them to a 2D segmentation mask. The researchers then use "region properties" to broadcast the updated material features and merge them into a 2D feature map. This feature map and the updated segmentation mask are used by a material generator to generate the final updated outfit. Before training, the system uses a classifier to learn "fashionable" images from fashion image data. These "fashionable" images serve as reference data for training. Researchers then obtain a large amount of clothing information from the network and gradually modify the images, replacing the fashion parts with the most dissimilar alternatives. This generates many negative samples to help AI learn what is considered unfashionable. Once the classifier is trained, the system gradually updates the outfits to make them more fashionable. The image generation neural network renders the adjusted appearance, using a variational autoencoder to generate contours and a conditional generative adversarial network (cGAN) to generate colors and patterns. Additionally, the latent encoding learned by the generator is used to identify which clothing items in the inventory can best achieve the desired outfit. In this study, the authors conducted experiments using the Chictopia10 dataset. They also used 15,930 images to train the generator and 12,744 images to train the fashion classifier. The dataset provides fashion images, while the remaining "unfashionable" images are modified by gradually replacing the fashion parts with different clothing items, as described earlier. Specifically, researchers selected a different outfit that is the farthest in Euclidean distance (using CNN-extracted latent vectors) from the replaced part to enhance the selection effect of "unfashionable" images. The construction of the training set is shown in the figure. The researchers modified some parts of the positive sample, replacing them with the least similar alternatives, resulting in negative samples (right image). The subtle changes in recommended outfits by Fashion++ are shown in the figure. Experiments have shown that the suggestions provided by this system make the images more realistic, and human evaluators find Fashion++'s suggestions both fashionable and easy to implement. Facebook believes that Fashion++ will help people learn fashion coordination through an application, without spending more money but by adjusting their existing clothing. Now, let's move on to shoes. A startup company called WANNABY, based in Belarus, has developed an app called "WANNA KICKS." This app combines augmented reality (AR) technology with smartphone cameras to allow users to see how shoes look on their feet. It can even simulate the various effects of shoes under different lighting conditions. Additionally, users can save and share AR virtual try-on photos on social media. The success of AR shoe try-on depends on the effectiveness of the software's recognition capabilities. Slow response, lag, or difficulty fitting the shoes can ruin the experience, regardless of how well the models are designed. During the experience with WANNA KICKS, there were no such issues. By following the prompts and placing your foot under the camera, the software quickly recognizes and generates an AR image of the shoes around your feet. The generated model fits well with your foot, whether you wear a size 42.5 like me or a size 46 like a friend with larger feet. Zooming in and out also results in immediate image transformations without any noticeable delay. This app quickly caught the attention of the Italian luxury brand GUCCI, and they collaborated with WANNABY to update their own app. This is undoubtedly a benefit for many young audiences who cannot afford GUCCI. If you can't buy it, why not give it a try? Fashion is a way of life, and everyone should have a passion for fashion. Each person is the protagonist of their own life, and walking down the street in carefully coordinated outfits is like walking on their own personal runway. The emergence of artificial intelligence is quietly changing our way of life and giving us a boost on our fashion journey. Embrace technology, embrace AI, and you will become a designer who controls the world of inspiration.

RECENT AI NEWS

RECENT AI TOOLS