Domain‐specific feature recalibration and alignment for multi‐source unsupervised domain adaptation

Traditional unsupervised domain adaptation (UDA) usually assumes that the source has labels and target no labels. In a real environment, labelled data comes from multiple different distributions. To handle this problem, multi-source (MUDA) is proposed. Multi-source aims to adapt model trained on multi-labelled domains unlabelled domain. paper, novel MUDA method by domain-specific feature recalibration alignment (FRA) Specifically, achieve recalibration, authors leverage channel attention pick out significant channels spatial focus important features in channels. Such integration of can lead effective may be great importance MUDA. addition, better MUDA, propose which consists Maximum Mean Discrepancy JS-divergence loss. reduce difference between Meanwhile, loss ensure prediction consistency classifiers domains. Four experiments have proved FRA significantly results popular benchmarks for