Domain Adaptive Semantic Segmentation without Source Data: 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) due to the large domain shift. 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, the accessing to 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 ATP, which consists of three schemes, ie, feature Alignment, bidirectional Teaching, and information Propagation. First, 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 are among the first ones to introduce negative pseudo labels to this field and develop a bidirectional self-training strategy to enhance the representation learning in the target domain. Finally, the information propagation scheme is employed to further reduce the intra-domain discrepancy within the target domain via pseudo semi-supervised learning. Extensive results on synthesis-to-real and cross-city driving datasets validate \textbf{ATP} yields state-of-the-art performance, even on par with methods that need access to source data.