Zero-Shot Learning with Partial Attributes

Abstract

Zero-shot classification, i.e. the task of learning predictors for classes with no training samples, requires to resort to subsidiary information in order to overcome the lack of annotated data. In the literature, one of the most popular approaches is to represent the class information by a set of visual attributes and to learn a visual-semantic embedding that allow us to transfer the information from those classes with plenty of annotated samples to those with no data available a training time. One mayor limitation of the attribute-based approach is that adding a new class requires a non-negligible annotation effort. This has motivated the search for alternative sources of semantic information. Here, the use of word embeddings learned from raw text appears as an appealing and scalable choice. In this paper, we consider a middleground scenario in which attribute vectors are only available for the training categories. We propose a deterministic approach to infer the attributes for the testing classes which, despite its simplicity, shows competitive results. We also propose two simple improvements to the structured embedding formulation of Akata et al., leading to significant improvements on attribute-only classification. Experiments on the Animals With Attributes and Caltech-UCSD Birds datasets show competitive performance.