Hierarchical domain adaptation with local feature patterns

• Main Authors: Wen, Jun, Yuan, Junsong, Zheng, Qian, Liu, Risheng, Gong, Zhefeng, Zheng, Nenggan
• 其他作者: School of Electrical and Electronic Engineering
• 格式: Article
• 語言: English
• 出版: 2023
• 主題: Engineering::Electrical and electronic engineering; Domain Adaptation; Local Feature Patterns
• 在線閱讀: https://hdl.handle.net/10356/164108
• 機構: Nanyang Technological University

Domain adaptation is proposed to generalize learning machines and address performance degradation of models that are trained from one specific source domain but applied to novel target domains. Existing domain adaptation methods focus on transferring holistic features whose discriminability is generally tailored to be source-specific and inferiorly generic to be transferable. As a result, standard domain adaptation on holistic features usually damages feature structures, especially local feature statistics, and deteriorates the learned discriminability. To alleviate this issue, we propose to transfer primitive local feature patterns, whose discriminability are shown to be inherently more sharable, and perform hierarchical feature adaptation. Concretely, we first learn a cluster of domain-shared local feature patterns and partition the feature space into cells. Local features are adaptively aggregated inside each cell to obtain cell features, which are further integrated into holistic features. To achieve fine-grained adaptations, we simultaneously perform alignment on local features, cell features and holistic features, within which process the local and cell features are aligned independently inside each cell to maintain the learned local structures and prevent negative transfer. Experimenting on typical one-to-one unsupervised domain adaptation for both image classification and action recognition tasks, partial domain adaptation, and domain-agnostic adaptation, we show that the proposed method achieves more reliable feature transfer by consistently outperforming state-of-the-art models and the learned domain-invariant features generalize well to novel domains.