Multi-level adversarial network for domain adaptive semantic segmentation

作者：

Highlights：

• Region-level adversarial learning enforces local region-level consistency across domains.

• Co-regularized adversarial learning ensures optimal multi-level alignment across domains.

• Multi-level consistency map guides domain adaptation in both input and output spaces.

• Proposed method outperforms the state-of-the-art with a large margin.

摘要

•Region-level adversarial learning enforces local region-level consistency across domains.•Co-regularized adversarial learning ensures optimal multi-level alignment across domains.•Multi-level consistency map guides domain adaptation in both input and output spaces.•Proposed method outperforms the state-of-the-art with a large margin.

论文关键词：Unsupervised domain adaptation,Semantic segmentation,Adversarial learning,Self training

论文评审过程：Received 4 January 2021, Revised 26 August 2021, Accepted 20 October 2021, Available online 22 October 2021, Version of Record 29 October 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.108384