RePatch: Region-Aware Patch Exchange for Cross-Domain Street Scene Segmentation

Abstract

To mitigate performance degradation caused by domain shifts, unsupervised domain adaptation (UDA) for semantic segmentation has been regarded as an important research direction, with potential applications for mobile robotics. However, global pseudo-label supervision is generally relied, and confidence variations are often overlooked across regions within the same image. To address this issue, a region-aware pseudo-label enhancement framework, termed RePatch, is proposed in the paper. Firstly, based on confidence and KL divergence, a bi-directional patch exchange mechanism is designed to achieve semantic complementarity between target and mixed domains. Then, dynamic semantic consistency replacement is conducted regarding low confidence regions. A weighting strategy for local confidence is designed to adaptively adjust contribution of different regions during loss optimization, thereby improving model robustness on uncertain areas. Comparative experiments demonstrate that the proposed framework achieves superior generalization ability and region discriminative ability in diverse cross-domain scenarios. In particular, local semantic consistency is obviously enhanced and structural boundary is effectively recognized. Note to Practitioners—This work is motivated by practical requirement for robust cross-domain semantic segmentation in autonomous driving and mobile robotics applications. Semantic segmentation models trained with labeled source-domain data often suffer performance degradation under variations in weather, illumination, viewpoints, and sensing conditions. Existing adaptation methods usually rely on global pseudo-label supervision, which may introduce unreliable supervision in uncertain regions, and thus reduce segmentation reliability in complex scenes. Through region-level semantic compensation and consistency constraints, the proposed RePatch framework improves pseudo-label reliability in low confidence regions, thereby enhancing robustness regarding structural boundaries and complex scenes. Experimental results demonstrate stable segmentation performance across multiple cross-domain scenarios, showing effectiveness of the proposed framework for autonomous perception under changing environments.