Continual coarse-to-fine domain adaptation in semantic segmentation
作者：
Highlights：
• We define a novel task combining coarse-to-fine learning and domain adaptation.
• Coarse-to-fine knowledge distillation transfers knowledge acquired on coarse classes.
• Coarse-to-fine unbiased weight initialization rule accounts for hierarchical splits.
• Maximum squares minimization addresses domain shift.
• We design and evaluate on two synthetic-to-real benchmarks in the driving scenario.
摘要
•We define a novel task combining coarse-to-fine learning and domain adaptation.•Coarse-to-fine knowledge distillation transfers knowledge acquired on coarse classes.•Coarse-to-fine unbiased weight initialization rule accounts for hierarchical splits.•Maximum squares minimization addresses domain shift.•We design and evaluate on two synthetic-to-real benchmarks in the driving scenario.
论文关键词：Coarse-to-fine learning,Unsupervised domain adaptation,Semantic segmentation,Continual learning,Deep learning
论文评审过程：Received 26 January 2022, Accepted 2 March 2022, Available online 8 March 2022, Version of Record 25 March 2022.
论文官网地址：https://doi.org/10.1016/j.imavis.2022.104426
