Abstract
Zero-Shot Learning (ZSL) seeks to recognize a sample from either seen orunseen domain by projecting the image data and semantic labels into a jointembedding space. However, most existing methods directly adapt a well-trainedprojection from one domain to another, thereby ignoring the serious biasproblem caused by domain differences. To address this issue, we propose a novelDomain-Specific Embedding Network (DSEN) that can apply specific projections todifferent domains for unbiased embedding, as well as several domainconstraints. In contrast to previous methods, the DSEN decomposes thedomain-shared projection function into one domain-invariant and twodomain-specific sub-functions to explore the similarities and differencesbetween two domains. To prevent the two specific projections from breaking thesemantic relationship, a semantic reconstruction constraint is proposed byapplying the same decoder function to them in a cycle consistency way.Furthermore, a domain division constraint is developed to directly penalize themargin between real and pseudo image features in respective seen and unseendomains, which can enlarge the inter-domain difference of visual features.Extensive experiments on four public benchmarks demonstrate the effectivenessof DSEN with an average of $9.2\%$ improvement in terms of harmonic mean. Thecode is available in \url{https://github.com/mboboGO/DSEN-for-GZSL}.