`Eyes of a Hawk and Ears of a Fox': Part Prototype Network for Generalized Zero-Shot Learning

Current approaches in Generalized Zero-Shot Learning (GZSL) are built uponbase models which consider only a single class attribute vector representationover the entire image. This is an oversimplification of the process of novelcategory recognition, where different regions of the image may have propertiesfrom different seen classes and thus have different predominant attributes.With this in mind, we take a fundamentally different approach: a pre-trainedVision-Language detector (VINVL) sensitive to attribute information is employedto efficiently obtain region features. A learned function maps the regionfeatures to region-specific attribute attention used to construct class partprototypes. We conduct experiments on a popular GZSL benchmark consisting ofthe CUB, SUN, and AWA2 datasets where our proposed Part Prototype Network (PPN)achieves promising results when compared with other popular base models.Corresponding ablation studies and analysis show that our approach is highlypractical and has a distinct advantage over global attribute attention whenlocalized proposals are available.