Source Data-Free Cross-Domain Semantic Segmentation: Align, Teach and Propagate

Benefiting from considerable pixel-level annotations collected from a specific situation (source), the trained semantic segmentation model performs quite well but fails in a new situation (target). To mitigate the domain gap, previous cross-domain semantic segmentation methods always assume the co-existence of source data and target data during domain alignment. However, accessing source data in the real scenario may raise privacy concerns and violate intellectual property. To tackle this problem, we focus on an interesting and challenging cross-domain semantic segmentation task where only the trained source model is provided to the target domain. Specifically, we propose a unified framework called \textbf{ATP}, which consists of three schemes, i.e., feature \textbf{A}lignment, bidirectional \textbf{T}eaching, and information \textbf{P}ropagation. First, considering explicit alignment is infeasible due to no source data, we devise a curriculum-style entropy minimization objective to implicitly align the target features with unseen source features via the provided source model. Second, besides positive pseudo labels in vanilla self-training, we introduce negative pseudo labels to this field and develop a bidirectional self-training strategy to enhance the representation learning in the target domain. It is the first work to use negative pseudo labels during self-training for domain adaptation. Finally, the information propagation scheme is employed to further reduce the intra-domain discrepancy within the target domain via pseudo-semi-supervised learning, which is the first step by providing a simple and effective post-process for the domain adaptation field. Furthermore, we also extend the proposed to the more challenging black-box source-model scenario where only the source model's prediction is available.Comment: 15 pages, 6 figure