Multi-Source Domain Adaptation with Collaborative Learning for Semantic Segmentation

• DOI: 10.48550/arxiv.2103.04717
• Publication Date: 2021-01-01
• Authors: Jianzhong He; Xu Jia; Shuaijun Chen; Jianzhuang Liu
• Keywords: Benchmark (surveying); Domain Adaptation

Multi-source unsupervised domain adaptation~(MSDA) aims at adapting models trained on multiple labeled source domains to an unlabeled target domain. In this paper, we propose a novel multi-source adaptation framework based collaborative learning for semantic segmentation. Firstly, simple image translation method is introduced align the pixel value distribution reduce gap between and some extent. Then, fully exploit essential information across domains, without seeing any data from addition, similar setting of adaptation, leveraged further improve performance adaptation. This achieved by additionally constraining outputs with pseudo labels online generated ensembled model. Extensive experiments ablation studies are conducted widely-used benchmark datasets in Our proposed achieves 59.0\% mIoU validation set Cityscapes training Synscapes GTA5 Cityscapes. It significantly outperforms all previous state-of-the-arts single-source methods.