Unsupervised Domain Adaptation for Semantic Segmentation Based on Instance Spatial Geometry

Unsupervised domain adaptation(UDA) for semantic segmentation transfers knowledge from synthetic to real domains, where geometric cues such as depth are commonly exploited to reduce the domain gap. However, existing depth-aware methods fail to explicitly model the domain-invariant spatial structures among semantic instances, and current self-training schemes weight pseudo-labels solely by prediction confidence, ignoring geometric consistency, which limits their reliability under domain shift. To address these issues, we propose an instance-wise geometric modeling framework that captures inter-instance spatial relations, including directional concentration and dominant direction, beyond conventional depth representations. These geometry-aware features are further integrated into pseudo-label weighting to enforce geometric-semantic consistency during self-training, leading to more stable and accurate pseudo supervision. Experiments on standard benchmarks show that our method significantly outperforms state-of-the-art UDA approaches.