Learning cross-domain semantic-visual relationships for transductive zero-shot learning

Zero-Shot Learning (ZSL) learns models for recognizing new classes. One of the main challenges in ZSL is domain discrepancy caused by category inconsistency between training and testing data. Domain adaptation most intuitive way to address this challenge. However, existing techniques cannot be directly applied into due disjoint label space source target domains. This work proposes Transferrable Semantic-Visual Relation (TSVR) approach towards transductive ZSL. TSVR redefines image recognition as predicting similarity/dissimilarity labels semantic-visual fusions consisting class attributes visual features. After above transformation, domains can have same space, which hence enables quantify discrepancy. For redefined problem, number similar pairs significantly smaller than that dissimilar ones. To end, we further propose use Domain-Specific Batch Normalization align