Diversely-Supervised Visual Product Search

This article strives for a diversely supervised visual product search, where queries specify diverse set of labels to search for. Where previous works have focused on representing attribute, instance, or category individually, we consider them together create visually describing products. We learn an embedding from the supervisory signal provided by every label encode their interrelationships. Once trained, has corresponding representation in space, which is aggregation selected items training set. At time, composite query representations retrieve images that match specific labels. form averaging over aggregated each For evaluation, extend existing datasets cars and clothes with Experiments show benefits our seen unseen combinations discovering design styles.