Facebook Gives You a Fashion Stylist, It Is Recommended to Check It Before Traveling on National Day
Scenario description:For many people, dressing is a difficult skill. Whether going to work, attending a meeting, dating, or going on vacation, everyone will be stumped by the question of "what to wear" in front of the wardrobe. Facebook has recently launched the AI fashion stylist Fashion++, which helps you make full use of existing clothes without having to buy a lot, and wear more fashionable effects, turning you into a fashion expert in seconds.
In the morning, I get up, wash up, put on beautiful makeup, and then click the "left" and "right" keys on the computer screen to watch the clothes on the screen switch back and forth. I pick a set of clothes I like and click "Dress Me". The screen will show the effect of the virtual portrait wearing the selected clothes. Today's outfit is easily prepared.
This is a clip from the movie Clueless. The fashionable heroine Cher Horowitz has a virtual stylist. Through this virtual stylist, she can preview the effects of her clothing combinations and choose a set of her favorite outfits.
Nowadays, the high-tech scenes in movies have been turned into reality by AI technology, and have even been surpassed.
Facebook The latest personalized clothing recommendation model Fashion++,Through algorithms, existing clothing can be cleverly adjusted to make them look more fashionable. Even a novice in dressing can instantly transform into a fashion expert.
AI Stylist: Dressing Skills at Your Fingertips
The most troubling questions for urban men and women every day are not only "What should I eat today?", but also "What should I wear today?"
Every time before going out, I would arrange the clothes in my closet in various combinations, but I still don’t know which combination looks best and is most suitable. So most female friends come to a conclusion: "There is always one piece of clothing missing in a woman's closet."
As the National Day holiday is approaching, are you going to travel and take beautiful photos? Do you feel that your wardrobe is lacking clothes? In fact, what you lack may not be clothes, but a dressing guide. In fact, the same piece of clothing can produce very different effects due to different ways of wearing it, such as adding a scarf or rolling up the sleeves.
Although today's major fashion bloggers with a large number of fans can help those who are not good at dressing up, they often end up just buying.
Fashion++, a fashion outfit recommendation model recently launched by Facebook, does not encourage shopping. Instead, it makes full use of users' existing clothing and adjusts the matching by changing the way they wear (such as rolling up sleeves or tucking them into trouser waists) to get the user's favorite look.
The Fashion++ model consists of a deep image-generating neural network that recognizes and learns to synthesize clothing conditioned on an encoding of each item of clothing. The encoding is explicitly decomposed in terms of clothing shape and texture, allowing for direct editing of fit and color/pattern/material, respectively.
The final experiment showed that Fashion++ was able toSuccessful outfit suggestions are provided and they are easy to implement.For example: get a new piece of clothing; adjust the color, the way it is worn (such as rolling up the sleeves), or the size (for example, making the pants looser).
How can AI systems be trained to provide dressing guidance?
French avant-garde fashion designer Coco Chanel once said, "Before you leave the house, look in the mirror and take one thing off." This quote has had a profound impact on the fashion world. The implication is that small adjustments, such as removing an accessory or taking off a hat, can make an existing garment more fashionable.
It was this idea that inspired the Fashion++ team to introduce a new computer vision challenge: hoping to improve the overall fashion sense by fine-tuning clothing/accessories through algorithms.
The specific contents of this project include: clothing generation framework, learning to judge fashion from online pictures, editing clothing combinations and outputting results.
Fashion++ Clothing Generation Framework
The text features t and shape features s of the initial garment x are initially edited and then edited by the editing module F++ to generate new text features and shape features t++ and s++.
Afterwards, the generator Gs The new features are fed back to the 2D semantic segmentation model and finally generated by the generator Gt Generate the edited new clothing matching effect x++.
Fashion++ Fashion Classifier
The team mapped the components of a given outfit (e.g., bag, top, boots) to their respective codes and then used a discriminative fashion classifier trained on 12,744 publicly available images of clothing that are recognized as fashionable.
In addition, the team also guided negative examples by swapping the clothing on the fashion examples with the most dissimilar clothing. When training the fashion classifier, the team considered that perhaps the ideal training set should consist of two groups of pictures: each group of pictures shows the same person wearing slightly different clothes, one of which is recognized as more fashionable than the other. However, they believe that such a dataset is not only unsuitable for large-scale curation, but also outdated as popular elements evolve.
Another approach is to consider a set of images from a certain group (such as celebrities) as positive examples, and another set of images as negative (such as everyday people). However, they found that such a dataset would lead to confusion between identity and style, so the classifier would find differences between the two groups on some attributes that have nothing to do with fashion.
Therefore, they proposed to automatically select less fashionable photos from online pictures of fashionable clothing. The main idea is to create "unfashionable" clothing from photos of fashionistas to train the model's recognition ability.
They started with a photo of a full outfit from Chictopia, an online fashion social network, selected one piece to modify, and then replaced it with a different piece. To increase the likelihood that the replacement piece would be less stylish, they took the piece from the set that was least similar to the original piece.
Using this data, the team trained a 3-layer Multilayer Perceptron (MLP) fashion classifier.By training a fashion classifier on these decompositions (to garments) and shape and texture encodings, a simple MLP can effectively capture subtle visual properties and complex garment synergies.
Edit the collocation and output the result
After the classifier is trained, the system will gradually update the clothing to make it more fashionable. 15930 imagesThe generator is trained.
The generator is composed of a neural network for image generation, a variational autoencoder for outline generation, and a conditional generative adversarial network (cGAN) for color and pattern generation, resulting in a newly adjusted appearance.
In order to take both pattern/color and shape/fit of clothing into account,They decomposed the encoding of each piece of clothing into texture and shape components.Allows the editor to control where and what to change (e.g., adjust the color of a shirt while keeping the style, or, change the neckline or tuck it in).
After optimizing the edits, the model provides its output in two formats, the first of which retrieves the garments from the inventory that best implement its suggestions, and the second of which plots how the same person would look in the adjusted appearance, based on the modified garment encoding.
The team validated its approach using Chictopia’s fashion image dataset and demonstrated through automated metrics and user studies that it can successfully generate minimal outfit edits while outperforming a baseline.
Facebook believes that Fashion++ will help people learn to dress fashionably using an app by tweaking existing clothing rather than spending more money on it.
Learn how to dress and be the coolest guy on the street
People have been pursuing fashion since ancient times, and how to dress more beautifully is an eternal topic. Because of this, we can see the continuous changes in clothing styles from ancient times to the present at home and abroad.
Some people think that fashion is art, some people think that fashion is happiness, no matter what, fashion is a symbol of beauty. Everyone wants to be fashionable, even to lead the fashion.
Nowadays, dressing fashionably is no longer just for the purpose of "women dress to please themselves", but also forExpression of self.
So, if you have already made your travel plans for the Golden Week, are you still wondering whether you should bring this scarf? Does this skirt match this shirt? How can you dress so as not to be lost in the crowd?
With Fashion++, a free personal clothing stylist, these problems should not be a problem. Hurry up and embrace technology and AI. Who knows, you may be the next fashion blogger.
-- over--