Towards Stable Source-Free Domain Adaptive Semantic Segmentation

We focus on Source-Free Domain Adaptation (SFDA) in semantic segmentation, which aims to adapt a source-trained model to a target domain without accessing source data. Self-training with a teacher-student framework is a common solution, but it often suffers from error accumulation caused by noisy pseudo-labels, leading to poor adaptability and unstable training. We analyze this issue and identify two key observations: (I) samples with high stability consistently achieve high accuracy, while unstable samples are the main source of errors; (II) stable self-training requires the teacher to incorporate improved students over time. Motivated by these insights, we propo‘se a novel Stable Self-Training (Stable-ST) framework, which consists of Stable Neighbor Denoising (SND) to correct pseudo-labels via bi-level neighbor-guided learning, and an enhanced Dynamic Teacher Update+ (DTU+) mechanism that leverages a new feedback metric, Neighbor Feature Density (NFD), to ensure reliable teacher updates. Furthermore, we provide a theoretical convergence analysis of SND, offering formal justification for its stability. Extensive experiments demonstrate that Stable-ST achieves state-of-the-art results across diverse SFDA settings, including traditional and continual urban adaptation, as well as medical and remote sensing scenarios. The source code is publicly available at https://github.com/DZhaoXd/Stable-ST.