Universal multi-Source domain adaptation for image classification
作者：
Highlights：
• The hypothesis of universal domain adaptation is violated in the multi-source setting.
• Existing multi-source domain adaptation methods primarily focus on the closed set setting.
• It is the target data that determines the common and private classes in the source domain.
• Samples in the same class should share a common weight during class-wise alignment.
• Model complexity should not increase with the change of domains.
摘要
•The hypothesis of universal domain adaptation is violated in the multi-source setting.•Existing multi-source domain adaptation methods primarily focus on the closed set setting.•It is the target data that determines the common and private classes in the source domain.•Samples in the same class should share a common weight during class-wise alignment.•Model complexity should not increase with the change of domains.
论文关键词：Universal domain adaptation,Multi-source domain adaptation,Universal multi-source domain adaptation,Universal multi-source adaptation network,Pseudo-margin vector
论文评审过程：Received 20 November 2020, Revised 30 June 2021, Accepted 6 August 2021, Available online 14 August 2021, Version of Record 19 August 2021.
论文官网地址：https://doi.org/10.1016/j.patcog.2021.108238
