Style-Hallucinated Dual Consistency Learning for Domain Generalized Semantic Segmentation

AbstractIn this paper, we study the task of synthetic-to-real domain generalized semantic segmentation, which aims to learn a model that is robust unseen real-world scenes using only synthetic data. The large shift between and data, including limited source environmental variations distribution gap significantly hinders performance on scenes. In work, propose Style-HAllucinated Dual consistEncy learning (SHADE) framework handle such shift. Specifically, SHADE constructed based two consistency constraints, Style Consistency (SC) Retrospection (RC). SC enriches situations encourages consistent representation across style-diversified samples. RC leverages knowledge prevent from overfitting data thus largely keeps models. Furthermore, present novel style hallucination module (SHM) generate samples are essential learning. SHM selects basis styles distribution, enabling dynamically diverse realistic during training. Experiments show our yields significant improvement outperforms state-of-the-art methods by 5.05% 8.35% average mIoU three datasets single- multi-source settings, respectively.