Semantic-diversity transfer network for generalized zero-shot learning via inner disagreement based OOD detector

Zero-shot learning (ZSL) aims to recognize objects from unseen classes, where the key is transfer knowledge seen classes by establishing appropriate mappings between visual and semantic features. Currently, in many existing works rather limited due various factors: (i) widely used features are global ones, they not completely consistent with attributes; (ii) only one mapping learned, which able effectively model diverse visual–semantic relations; (iii) bias problem generalized ZSL (GZSL) could be handled. In this paper, we propose two techniques alleviate these limitations. Firstly, a Semantic-diversity Network (SetNet) addressing first limitations, (1) multiple-attention architecture diversity regularizer proposed learn multiple local being more attributes (2) projector ensemble geometrically takes as inputs diversify relations. Secondly, an inner disagreement based domain detection module (ID3M) for GZSL problem, picks out unseen-class data before class-level classification. Due lack of training stage, ID3M employs novel self-contained scheme detects on criterion. Experimental results three public datasets show that SetNet explored achieves significant improvement against state-of-the-art methods.