Adversarial Self-Training with Domain Mask for Semantic Segmentation

Unsupervised Domain Adaptation (UDA) faces challenges in real-world semantic segmentation tasks due to the large difference between labeled training (source) data and unseen test (target) data, especially in driving assistance and autonomous driving scenarios. In particular, this large domain gap may cause a significant performance drop. Various adversarial learning methods were proposed to reduce the domain variance in the feature space and output space (i.e., the output feature space of the segmentation network). However, the domain invariance knowledge is seldom taken into consideration. In this paper, we proposed an Adversarial Self-Training (AST) framework with the observation of the generated pseudo-labels holding the class-balanced intrinsic domain invariance that can be integrated into both output-space and adversarial-space (i.e., the output space of the discriminator network). Furthermore, we separate the modulation of Batch Normalization (BN) layers from training, in order to concentrate the task of adversarial learning and elevate the quality of pseudo-labels. The experiment results show the proposed methods perform favorably against the state-of-the-art methods in terms of accuracy.