search
top

Facebook Project Fashion++: Minimal Edits for Outfit Improvement

In a latest research by Facebook, Machine Learning models are designed to do tasks to mimic a fashion advisor. Such tasks being carried out by humans is quite common, but seems extraordinarily difficult to be achieved by a computer.

 
 
 
The research was presented at the International Conference on Computer Vision 2019 held during October 27th – November 2nd, 2019 at Seoul, Korea. This was one of the papers among the list of research papers presented in the conference specifically on Artificial Intelligence in the area of Computer Vision. This is clearly evident from the research work presented at the conference that the research at Facebook is rigorously working to find what all can be achieved through Computer Vision.

 
Facebook research team have introduced an image generation framework to do minimal edits on outfits making them more fashionable.After successfully building the application, some people were asked to evaluate Fashion++ by showing them a pair of images at the same time; one of original outfit and the other one edited by Fashion++. The results are quite promising.

 

The Project

The project is named Fashion++ , which is based on a large library of images labeled with both the pieces worn (e.g. hat, scarf, skirt) taking into account the overall fashionability of a subject. It consists of features which suggest some minor changes in a given outfit. Although, the researchers do not call it a complete digital fashion assistant, but they observed success in giving suggestions for outfits for people.

 
As the project is aimed at suggesting minimal edits in the outfit of a person, it makes use of baselines like:

  • Similarity only: This tries to maintain the similarity of new outfit with the current by selecting the nearest neighbor garment. This allows to make least changes in the outfit, but it does not improve fashionability.
  • Fashion only: This changes to the outfit to another one which gives the highest fashionability score predicted by the classifier. This tries to improve the fashionability score the most, changing the outfit significantly.
  • Random sampling: This changes to a randomly sampled garment. This neither improves fashionability nor remains similar.

https://research.fb.com/publications/fashion-minimal-edits-for-outfit-improvement/(Courtesy: Research paper on Facebook)

 

How it Happened?

Training a network to suggest changes in the outfit requires a lot of data. A dataset consisting of 15,930 images is used to train the generators while about 12,744 images are used for training fashionability classifier. With the help of clearly defined procedure a set of positive and negative examples are then formed for training the fashionability classifier. Once trained, more than 3,000 examples are used to evaluate the performance of classifier. The test examples are chosen to match those from real world outfits which makes it possible to test the classifier on real data.

A good outfit generator module is required to obtain an accurate fashionability classifier with required amount of editing. Failing to achieve any of these may end up getting a worst outfit changer.

Performing minimal edits in an outfit is name sake. As such there are various possibilities within 2 broad options: changing shape of outfit or changing the texture of outfit. Changing shape may require to change the length, fitting, adjusting waistline etc. while changing texture may need to swap the colors, making it monochrome or polishing the outfits.

Pointing out redundant pieces of outfits and suggesting new ones which go well with those which are there can be tricky. This can significantly change the looks of the person wearing the outfit.

 

Applications

If you are still thinking what are the possible applications of this project, then keep reading.

  • Artificial Intelligence has acquired a strong place in the area of animations and virtual environments. This has opened ideas for animators who would like to match it to their animated characters in real life. This can help them to select the best outfit to suit their animated characters.
    One of the applications can be to make someone look more fashionable. This may bring confidence in them by wearing outfits that suit them.
  • A more interesting application can be looking in the mirror and helping you look better by advising you some small alterations in your cloths.

If a system can give you suggestions in improving your outfit, what’s better than that.

One Response to “Facebook Project Fashion++: Minimal Edits for Outfit Improvement”

  1. Bina Vani says:

    Facebook is doing great job in research and so you are.

Leave a Reply

Your email address will not be published.

top