Divide and augment: Supervised domain adaptation via sample-wise feature fusion

The training of deep models relies on appropriate regularization from a copious amount of labeled data. And yet, obtaining a large and well-annotated dataset is costly. Thus, supervised domain adaptation (SDA) becomes attractive, especially when it aims to regularize these networks for a data-scarce target domain by exploiting an available data-rich source domain. Different from previous methods focusing on an cumbersome adversarial learning manner, we assume that a source or target sample in the feature space can be regarded as a combination of (1) domain-oriented features (i.e., those reflecting the difference among domains) and (2) class-specific features (i.e., those inherently defining a specific class). By exploiting this, we present Divide and Augment (DivAug), a feature fusion-based data augmentation framework that performs target domain augmentation by transforming source samples into the target domain in an energy-efficient manner. Specifically, with a novel semantic inconsistency loss based on a multi-task ensemble learning scheme, DivAug enforces two encoders to learn the decomposed domain-oriented and class-specific features, respectively. Furthermore, we propose a simple sample-wise feature fusion rule that transforms source samples into target domain by combining class-specific features from a source sample and domain-oriented features from a target sample. Extensive experiments demonstrate that our method outperforms the current state-of-the-art methods across various datasets in SDA settings.