Class-attribute inconsistency learning for novelty detection

•A new notion of class-attribute inconsistency for novelty detection.•A novelty often has inconsistent class- and attribute-level similar references.•CAILNet outperforms state-of-the-arts by exploring and mining inconsistency. In this paper, we address the problem of novelty detection whose goal is to recognize instances from unseen classes during testing. Our key idea is to leverage the inconsistency between class similarity and (latent) attribute similarity. We are motivated by the observation that a novel class may holistically appear like a certain known class (class-level reference) but often exhibits unique properties similar to others (attribute-level references). That is, the related class- and attribute-level references are often inconsistent for a novel class. A new two-stage Class-Attribute Inconsistency Learning network (CAILNet) is proposed to explore class-attribute inconsistency for novelty detection. Stage one aims to learn both class and attribute features based on the class labels and fake attribute labels, and stage two aims to search for the corresponding references and make fine-grained comparisons for final novelty decision. Empirically we conduct comprehensive experiments on three benchmark datasets, and demonstrate state-of-the-art performance.