Learning discriminative instance attribute for zero-shot classification

We study the problem of zero-shot classification in which we don't have labeled data in target domain. Existing approaches learn a model from source domain and apply it without adaptation to target domain, which is prone to domain shift problem. To solve the problem, we propose a novel Learning Discriminative Instance Attribute(LDIA) method. Specifically, we learn a projection matrix for both the source and target domain jointly and also use prototype in the attribute space to regularise the learned projection matrix. Therefore, the information of the source domain can be effectively transferred to the target domain. The experimental results on two benchmark datasets demonstrate that the proposed LDIA method exceeds competitive approaches for zero-shot classification task.