Descripció del projecte
Social Media has a great impact on Fashion and the multi-billion dollar fashion industry. Online sales in the year 2017 are expected to reach 191 billion Euros in Europe and 370 billion dollars in the USA. Providing insights about fashion trends from social media and fashion blogs improves the interaction within fashion industries as well as between sellers and buyers significantly. However, this task meets various difficulties due to the large amount of raw data and noisy information obtained from internet.
Manually supervising these meta-data as well as the abundant photos existing on numerous social networks like Instagram, etc, are extremely costly. However, recent breakthroughs in Computer Vision shot that it is possible to use such meta information as weak labels for Semi-supervised or even Unsupervised learning strategies. It has also been shown that multi-objective models (obtained by training multiple objectives at the same time) often perform better than single-objective models. Yet, in current approaches no objective possesses information about the others. This is where the application of Structured Learning techniques becomes helpful: we can improve the model performance by considering the interrelations between all the objectives in the learning process. For instance, such structured models could be used to gather fashion-related information, such as color or pattern, from fashion blogs, social media or e-commerce sites to gain insights regarding current fashion styles or trends. In this line of work, we propose to design novel deep learning algorithms for visual understanding using weak data obtained from web. This novel method will facilitate the interaction between different players in the fashion industry, including the interaction between departments of a single fashion corporation, and also between companies and consumers.
The upsurge of computer vision, GPU computation and the breakthroughs in the field of Deep Learning, has opened a distinct line of research and applications. Over this course, Computer Vision and Computational Learning Group (CVCL) of the University of Barcelona as a main branch of UB Data Science team aims to combine synergies with a view to undertaking technology transfer projects in the field of data science. CVCL team has been involved in various industrial research projects in the field of data science such as model for estimating the click-through rate of an ad, diagnostic system for intestinal motility based on medical images, aesthetic image assessment, etc.
