Self-supervised Augmentation Consistency for Adapting Semantic Segmentation

We propose an approach to domain adaptation for semantic segmentation that is\nboth practical and highly accurate. In contrast to previous work, we abandon\nthe use of computationally involved adversarial objectives, network ensembles\nand style transfer. Instead, we employ standard data augmentation techniques\n$-$ photometric noise, flipping and scaling $-$ and ensure consistency of the\nsemantic predictions across these image transformations. We develop this\nprinciple in a lightweight self-supervised framework trained on co-evolving\npseudo labels without the need for cumbersome extra training rounds. Simple in\ntraining from a practitioner's standpoint, our approach is remarkably\neffective. We achieve significant improvements of the state-of-the-art\nsegmentation accuracy after adaptation, consistent both across different\nchoices of the backbone architecture and adaptation scenarios.\n