Descripció del projecte
Fashion has a large impact on society and is a multi-billion dollar industry. Online sales in the year 2017 reached 191 billion euros in Europe and 370 billion dollars in the USA. Developing tools to improve the interaction between sellers and buyers will be of paramount importance.
One such tool is the ability to automatically determine all the attributes and characteristics (facets) of a fashion product from a single picture. This level of understanding would allow improvements in several areas, like more dense and fine-grained text-based product search, more nuanced understanding of trends and styles, better synthesis of natural language related to fashion products, or better image representations for fashion products, useful for visual search or visual question answering. All of them, technologies that could be deployed in useful products for the online fashion retail industry.
However, attributes in the fashion industry are difficult to enumerate precisely, as they are constantly changing and subject to trends. Moreover, fashion is a notoriously hard domain for fine-grained classification, with blurry and ill-defined boundaries between categories. Therefore our objective is to discover a dynamic taxonomy of fashion concepts and attributes by analyzing a large corpus of multi-modal fashion related data, and subsequently learn to detect them in images. This taxonomy will evolve with time, to keep up with new styles and trends.
In order to train a system capable of automatically reasoning about fashion products, the ability to leverage large amounts of fashion data will be paramount. Processing this “big data” will involve the development of advanced multi-modal machine learning algorithms, and making them scalable to large volumes of data. Likewise, discovering the relevant information, gathering it from online sources and organizing it into coherent datasets will be part of the work. The current revolution in machine learning, propelled by big data, GPU computation and new theoretical breakthroughs (especially Deep Learning), opens up many new applications which were previously unattainable.
The SGR Intelligent Data Science and Artificial Intelligence (IDEAI-UPC) devotes its research efforts to the resolution of real problems, in the immediate economic and industrial context. It is the result of the fusion of different groups of the UPC, all with more than 30 years of history and many scientific contributions. Currently the group focuses its efforts in the areas of Machine Learning, Data Science, Natural Language Processing, Intelligent Decision Support Systems, Cognitive systems and Applications, and is one of the leading groups in Spain in these areas. This industrial project would allow Wide Eyes Technologies to directly apply the latest scientific advances in the field to develop innovative fashion applications.
Regarding the project planning, in a first stage the student will perform an in depth study of available machine learning techniques for natural language processing and computer vision. Also in this phase the student will work on data acquisition, and tools for automatic Internet crawling and exploration of new meta-data which can be automatically extracted from e-commerce sites, blogs or social media. The second phase will focus on developing algorithms for efficient attribute discovery and learning in fashion data, keeping in sight the scalability with data set size, and computational cost that is essential for commercial deployment. At all stages Wide Eyes Technologies is interested in protecting the intellectual rights by means of patenting, or publishing results in international conferences.
