Structured discriminative tensor dictionary learning for unsupervised domain adaptation

Unsupervised domain adaptation aims at learning a classification model robust to data distribution shift between labeled source and an unlabeled target domain. Most existing approaches have overlooked the multi-dimensional nature of visual data, building models in vector space. Meanwhile, issue limited training samples is rarely considered by previous methods, yet it ubiquitous practical applications. In this paper, we develop structured discriminative tensor dictionary method (SDTDL), which enables matching SDTDL produces disentangled transferable representations explicitly separating domain-specific factor class-specific data. Classification achieved based on sample reconstruction fidelity alignment, seamlessly integrated into learning. We evaluate cross-domain object digit recognition tasks, paying special attention scenarios test beyond set. Experimental results show that our outperforms mainstream shallow representative deep methods significant margin.