AI-enabled programs can think, learn, and act in particular situations as humans do. When these intelligent programs are applied to machines, they can understand and improve without human intervention. E-commerce is one such everyday use case where merchants are increasing their sales by leaps and bounds using this technology. But AI-powered reposing systems for ecommerce use cases have its features and potential shortcomings. Consider them using an example Goodstyle app — a multi-brand virtual fitting room.
Algorithms of Goodstyle App are implemented using a neural network. With the help of the service (App Store / Play Market), the user can not only try on things virtually, but also combine items, create a complete look and manage a virtual wardrobe.
Goodstyle uses neural networks to quickly process and add clothes to the catalog. Another neural network is in the process of development, it will make the application more personalized. Not only parameters will be taken into account, but also the preferences of each specific user. The program will calculate and recommend images by itself. This is a new level, the level of a virtual stylist, when it is not a person who advises, but a neural network.
How were systems trained
The developers of GoodStyle App use various methods to detect clothes. At the moment the models of neural networks let classifying several dozens of clothing categories, applying the segmenting mask and detecting key points on almost every piece of clothing. Developers learn how to transfer pieces of clothing from one image to another, to transform it for virtual trying on on avatars.
They use different types of neural networks and libraries in research (including mask rcnn, cyclegan, stylegan2 and opencv) and choose the best ones in accordance with the results they get. In order to get the optimal result they try to use several networks at a time. Anyway in all our experiments they work with clothing images, shapes, textures, colors, in other words, with an array of pixels of images. This is why their results are high-qualified and appeal to any body type. Obviously, trying to solve the problem of virtual trying on, impossible to process all the existing body types people have. This is why the number of body types is now limited within GoodStyle App. They are doing research using GAN nowadays that will help users to experience virtual trying on using their personalized and detailed body type if the given prototypes do not match.
What are the limitations of techniques like GANs, both technical and practical?
GAN research is one of the latest updates in development of Goodstyle app. Neural networks of this type are highly ranked by developers due to the received results and potential opportunities. New types of GAN are presented within a short period of time while the existing ones are being improved and finished up. Things that were practically impossible to realise using classical neural networks are getting real with GAN.
GAN research is a seperate way of development in company. With the help of this solution developers improve the solutions based on basic neural networks. In particular the cases of transforming style, shape, color from one picture to another. The cases of virtual trying on for any body type are expected to be solved with GAN too. Technically there are no limits to this solution nowadays. Developers use Amazon cloud services that offer practically unlimited volumes of graphical processes. Thanks to them they are able to train the network to solve any cases in a short period of time. As for technical limits, on this way of development networks open new horizons fot them and they are super excited to use them, seeing no boundaries for this process.
Nevertheless there are several problems:
- time on training and hardware requirements. GAN requires much time on training. Using one type of GPU may take one period of time on training, while using another type of GPU may take much more time on the same purpose. This is why iterations of networks’ development and training (there are usually dozens of them, if not hundreds) are much longer than typical for our developers two-week periods.
- Each part of GAN is able to overcome the other one. If the discriminator is too good then it will return values very close to 0 or 1 so that the generator will have troubles reading the gradient. If the generator is too good then it will be constantly using discriminator’s disadvantages, leading to incorrect results.
Benefits of AI-powered reposing systems for e-commerce
AI-driven personalization helps retailers:
- Increase conversion rates across all channels.
- Improve the overall user experience.
- Maximize ROI.
What would the Goodstyle company say to someone who argues that its systems could put clothing models out of work? Currently team is developing a solution that will complement the existing selling methods based on presenting clothes on models. They are not going to exclude models out of the selling process as the pictures of the models that they take are the basis for content. The context of business-case is the statement that pictures of clothes do not exist without a model, this is why solution has to be “smart” enough to convert the original image of the model to 2D (and to 3D in future).
The main aim of app is to help users find their own personal style without even leaving their place. Using GoodStyle App any person can virtually try on various outfits, doing practically nothing, just using a smartphone, and to order clothing online or to visit an offline store, being 100% sure what to buy there. This is how they can directly affect the selling process, increasing sales and stimulating vendors to release new collections. The more new collections – the more work for models! It is obvious that in this case they work on their side.