Hierarchical domain adaptation with local feature patterns

Highlights：

• We propose to learn primitive local feature patterns for unsupervised domain adaptation, whose discriminability is inherently more transferable than the typically-adopted holistic features.

• We propose a hierarchical feature adaptation strategy to achieve fine-grained feature alignments, through which discriminative local feature structures are maintained and negative transfer of irrelevant features is attenuated.

• Experimenting on the typical one-to-one domain adaptation for image classification and action recognition tasks, challenging partial domain adaptation and domain-agnostic learning, the consistently distinct improvements demonstrate the superiority of the proposed method over state-of-the-art approaches. Further, our learned domain-invariant features are shown to generalize well to novel domains.