Student: Francesca Del Lungo, 2021. This thesis was developed during a stage at Intuendi srl, Florence
Clothing classification algorithms often face several challenges. First of all,
clothing items often have many variations in style, texture and cutting. Second, clothing attributes are fine-grained properties that belong to a specific part of the image. Two main approaches are developed to predict visual attributes from the images: the first treats the problem as a multi-label classification and the other as a multi-task multi-label classification. We obtain very good results outperforming the state-of-the-art models, even those methods that need landmarks, which are not required in our approach. We achieved 70% top-5 recall and 60% top-3 recall on both texture and shape-related attributes and 55% top-5 recall on distinctive
clothes attributes, like sleeve-length and neckline type. The case study underlying this work carried out in collaboration with the Intuendi ompany, has shown that this attribute extraction technique can also bring great benefits to products demand forecasting, especially when historical sales data is not available or is not enough to predict future demand in a robust way.