Distribution regularized self-supervised learning for domain adaptation of semantic segmentation

This paper proposes a novel pixel-level distribution regularization scheme (DRSL) for self-supervised domain adaptation of semantic segmentation. In typical setting, the classification loss forces segmentation model to greedily learn representations that capture inter-class variations in order determine decision (class) boundary. Due domain-shift, this boundary is unaligned target domain, resulting noisy pseudo labels adversely affecting adaptation. To overcome limitation, along with capturing variation, we intra-class through class-aware multi-modal learning (MMDL). Thus, information necessary explicitly disentangled from discrimination. Features captured thus are much more informative, pseudo-labels low noise. disentanglement allows us perform separate alignments discriminative space and space, using cross-entropy based self-learning former. For later, propose stochastic mode alignment method, by decreasing distance between source pixels map same mode. The metric loss, computed over backpropagated modeling head, acts as regularizer base network shared head. results comprehensive experiments on synthetic real setups, i.e., GTA-V/SYNTHIA Cityscapes, show DRSL outperforms many existing approaches (a minimum margin 2.3% 2.5% mIoU SYNTHIA Cityscapes).