Source-Free Open Compound Domain Adaptation in Semantic Segmentation

In this work, we introduce a new concept, named source-free open compound domain adaptation (SF-OCDA), and study it in semantic segmentation. SF-OCDA is more challenging than the traditional but practical. It jointly considers (1) issues of data privacy storage (2) scenario multiple target domains unseen domains. SF-OCDA, only source pre-trained model are available to learn model. The evaluated on samples from To solve problem, present an effective framework by separating training process into two stages: pre-training generalized adapting with self-supervised learning. our framework, propose Cross-Patch Style Swap (CPSS) diversify various patch styles feature-level, which can benefit both stages. First, CPSS significantly improve generalization ability model, providing accurate pseudo-labels for latter stage. Second, reduce influence noisy also avoid overfitting during learning, consistently boosting performance Experiments demonstrate that method produces state-of-the-art results C-Driving dataset. Furthermore, achieves leading CityScapes generalization.