FashionNet: Recognition of fashion styles in the wild

During the last years fashion has experienced a great evolution. New technologies and access to the Internet have revolutionized such a mature market. The manufactuting of products, the way to advertise them, and their life cicles (more ephemeral) have changed. In addition, the target audience, the distribution channels, the shopping experience are not the same anymore. Due to these facts, new business models and new opportunities are emerging.

Nowadays, there is a growing interest in the development vision systems to recognize, classify, and recommend fashion styles to users, in order to learn their personal styles preferences and recommend new products to them. Machine learning techniques have the potential of building these functionalities to create different services and applications that adapt to the new scenario of fashion.

Recent advances in the recognition of clothing have been motivated by the collection of large clothing datasets and algorithms for their treatment. However, existing databases are limited in the amount of annotations, and not realistic to address the various challenges that present real applications, such as the high variability and deformability in the appearance of clothes, model poses, viewpoints, etc.

This project presents a clothing-style recognition system based on neural networks. To that end, a dataset with different clothing styles has been collected. This system is a supervised learning system based on Deep Learning technology through convolutional neural networks, which are capable of to classify the different patterns of clothing styles through images taken in-the-wild. As a result, the system is able to offer a more direct and personalized experience to users. Unlike other works focused on hand-crafted features or descriptors, the proposed system presents a higher generalization power, and achieve a higher recognition rate.

The development of such system can encourage future research to develop new recognition techniques for fashion with the aim of minimizing the time of searching specific clothes, and improve the user experience while shopping.


Ana I. Maqueda:

Carlos R. del Blanco Adán: