Ben-Cohen Avi ; Zamir Nadav ; Baruch Emanuel Ben ; Friedman Itamar ; Zelnik-Manor Lihi

Abstract

Training a neural network model for recognizing multiple labels associatedwith an image, including identifying unseen labels, is challenging, especiallyfor images that portray numerous semantically diverse labels. As challenging asthis task is, it is an essential task to tackle since it represents manyreal-world cases, such as image retrieval of natural images. We argue thatusing a single embedding vector to represent an image, as commonly practiced,is not sufficient to rank both relevant seen and unseen labels accurately. Thisstudy introduces an end-to-end model training for multi-label zero-shotlearning that supports semantic diversity of the images and labels. We proposeto use an embedding matrix having principal embedding vectors trained using atailored loss function. In addition, during training, we suggest up-weightingin the loss function image samples presenting higher semantic diversity toencourage the diversity of the embedding matrix. Extensive experiments showthat our proposed method improves the zero-shot model's quality in tag-basedimage retrieval achieving SoTA results on several common datasets (NUS-Wide,COCO, Open Images).