VIRUSA- A Virtual Trial room Assistant
INTRODUCTION
Trial rooms are used by shoppers to try-out articles of clothing that they intend to purchase. In an age, where consumer shopping has started to become largely online, driven by increased internet access, comfort and convenience, it has become increasingly important for online retailers to provide a means by which shoppers can try the goods that they intend to purchase. This sentiment has led us to create a solution -VIRUSA which makes use of various deep learning, image processing techniques to indicate whether a particular apparel fits on your body or not. The application allows the user to feed in an image of their body and an image of the clothing which they need to try on, and provides the results in the form of image of their body wearing that desired outfit.
WHAT IS THE PROBLEM?
Currently, there are a few major issues that customers face while shopping for clothes. For those who shop at retail outlets, suffer from the inconvenience of waiting in long queues for trial rooms and are even sometimes restricted on the number of articles that they can carry into the trail room at a time. Not only this, many a times articles they might like become unavailable in their required size due to inventory restrictions, adding greatly to their frustrations and making for a tiring shopping experience. On the other hand, those who shop online to fulfill their clothing needs are often dissatisfied when the clothes in reality don’t match with their perception derived from pictures displayed on website or when the clothes aren’t the right fit for them, compelling them to return the product or seek a replacement. Looking at the same situation from the perspective of online retailers this translates to the loss of a very large customer base of about 80% of the online shoppers. Further the costs of last mile delivery which account for 53% of overall logistic costs, makes the system of returns and replacements a cause of concern for online retailers.
NEEDS
For the section of online customers, there is a need to help them better perceive what their article of choice looks like in reality, while for online retailers, need for a system to reassure their customers into buying their products. This product will be geared towards enabling end users, shoppers, to make educated shopping decisions and identify the right fit and look of outfit that they seek to buy without physically trying them out adding to their comfort and convenience. This enhanced shopping experience will improve customer satisfaction, increasing the number of buyers, and hence making a larger target market for online retailers. Moreover, the reduction in number of returns and replacements would translate to reduced last mile delivery costs for them.
PROJECT APPROACH
Virusa allows users to feed in the image of a user and an article of clothing, which it then uses to render an image of the user wearing that article of clothing. CNNs were used to extract features from the images of the person and the article of clothing which were then combined into a single tensor using a correlation layer. The tensor was then passed to a regression network that predicted the spatial transformation parameters. A TPS module was then used to warp the article of clothing to the pose of the user. UNet was then used to render the image of the person wearing the article of clothing and the composition mask. Finally the rendered image of the user and the warped article of were combined using a composition mask.
Towards the development of our project, we developed the front-end using HTML, CSS, Bootstrap. Flask was being used for web server. For the purpose of managing user registration and authentication module, MongoDB Atlas cluster was used. Pytorch along with CUDA toolkit was used to train rendering model on Nvidia GPU. We explored various possible approaches for our objective, did an extensive literature survey to choose the best approach, then compared and finalized best suitable models to be used for respective modules in the VIRUSA pipeline.
CONCLUSION
VIRUSA makes use of deep learning, image processing techniques and various models to give the user a hint of look and fitting of the apparel they like, without having to actually try them physically. It saves the user a lot of time and from the discomfort caused by having to wait for vacant trial rooms in shopping stores and from the hassle of changing clothes. It also eliminates one’s need of visiting clothing stores as one can shop online and still get the feel of clothing on them while sitting at their homes. The application doesn’t require any specialized and costly hardware set-ups and can be used on handheld devices like phones or tablets and allows users to check the look and fit of articles of clothing or accessories on them virtually anywhere.
