Multi-teacher Self-training and Semantic-augmented Alignment for Multi-Target Domain Adaptive Semantic Segmentation

Multi-target domain adaptive semantic segmentation (MTDA) trains a single model on labeled data from one source domain to generalize across multiple unlabeled target domains. Existing MTDA methods typically train multiple domain-specific models and then distill their knowledge into a unified model. In this paper, we propose a novel multi-teacher self-training framework for MTDA by introducing a mean teacher (MT) model, which consists of a shared feature extractor and domain-specific classifiers. For each target domain, the teacher classifier, constructed by Exponential Moving Average of the historical versions of the student, distills its knowledge to the current student classifier. Moreover, we design a domain-agnostic classifier that learns jointly from all domain-specific teacher classifiers. Leveraging the MT architecture, our design enhances the consistency of self-training across multiple target domains. Additionally, to mitigate the class-mismatch issue, we propose semantic-augmented feature alignment, which aligns target features with a dynamically updated source feature dictionary during training. Extensive experiments on multiple benchmarks demonstrate that our method achieves state-of-the-art performance in MTDA tasks. The code is available athttps://github.com/qh-ujs/SAFA.