Towards self-similarity consistency and feature discrimination for unsupervised domain adaptation

作者：

Highlights：

• A novel self-similarity consistency (SSC) metric is proposed to measure the domain discrepancy.

• We demonstrate that GMM and CORAL can be viewed as a special case and a sub-optimal measure of the proposed SSC metric.

• We propose a simple yet effective approach to enlarge the separability of inter-class samples, which improves the adaptation performance significantly.

摘要

•A novel self-similarity consistency (SSC) metric is proposed to measure the domain discrepancy.•We demonstrate that GMM and CORAL can be viewed as a special case and a sub-optimal measure of the proposed SSC metric.•We propose a simple yet effective approach to enlarge the separability of inter-class samples, which improves the adaptation performance significantly.

论文关键词：Domain adaptation,Self-similarity consistency,Feature discrimination,Intra-class compactness,Inter-class separability

论文评审过程：Received 3 April 2020, Revised 14 October 2020, Accepted 4 March 2021, Available online 10 March 2021, Version of Record 16 March 2021.

论文官网地址：https://doi.org/10.1016/j.image.2021.116232