Automatic Check-out within Generalized Zero-shot Learning Setting

Automatic Check-Out (ACO) aims to accurately forecast the presence and counts of each product in any given product combination. This paper concentrates on the ACO problem within generalized zero-shot Learning (GZSL) setting. Many recent GZSL studies have adopted a feature generation approach. However, these methods treat each class synthesis process independently, ignoring the rich dependency structure of different classes and their compositions. We propose a generative model that captures the common structure between classes and a classifier which optimizes classification boundaries by imposing a multi-class Deep SVDD constraint on seen classes. The evaluations are performed on the large-scale RPC dataset during the experiments.