Distribution regularized self-supervised learning for domain adaptation of semantic segmentation

作者：

Highlights：

• We propose a novel distribution regularization scheme (DRSL) for self-supervised domain adaptation of semantic segmentation.

• Along with capturing inter-class variations, we capture pixel-level intra-class variations through class-aware MMDL-FR.

• Separate alignments are performed in discriminative space (self-training) and multi-modal distribution space (MMDL-FR).

• We show that due to the regularization of MMDL-FR, the pseudo-labels generated over the target domain are more accurate.

• We present state-of-the-art performance for benchmark synthetic to real, e.g., GTA-V/SYNTHIA to Cityscapes adaptation.

摘要

•We propose a novel distribution regularization scheme (DRSL) for self-supervised domain adaptation of semantic segmentation.•Along with capturing inter-class variations, we capture pixel-level intra-class variations through class-aware MMDL-FR.•Separate alignments are performed in discriminative space (self-training) and multi-modal distribution space (MMDL-FR).•We show that due to the regularization of MMDL-FR, the pseudo-labels generated over the target domain are more accurate.•We present state-of-the-art performance for benchmark synthetic to real, e.g., GTA-V/SYNTHIA to Cityscapes adaptation.

论文关键词：Semantic segmentation,Self-supervised learning,Domain adaptation,Multi-modal distribution learning

论文评审过程：Received 12 March 2022, Revised 2 June 2022, Accepted 6 June 2022, Available online 13 June 2022, Version of Record 8 July 2022.

论文官网地址：https://doi.org/10.1016/j.imavis.2022.104504