Unsupervised Domain Adaptation to Improve Image Segmentation Quality Both in the Source and Target Domain

Domain adaptation is becoming more and more important with the advancing development of machine learning and the ever-increasing diversity of available data. The advancement of autonomous driving depends very much on progress in machine learning, which relies heavily on vast amounts of training data. It is well known that the performance of such models drops, as soon as the data used during inference stems from a different domain as the training data. To avoid the need to label a separate dataset for each new domain, e.g., each new camera sensor, methods for domain adaptation are necessary. Most interesting are unsupervised domain adaptation approaches since they do not require costly labels for the target domain. In this paper we adapt a known domain adaptation approach to work in an unsupervised fashion for semantic segmentation on high resolution data and provide some analysis of the learned representations. With our domain-adapted semantic segmentation we were able to achieve a significant 15 % absolute increase in mean intersection over union (mIoU), securing a surprisingly good 5th rank on the target domain Kill I test set without having used any Kill I labels during training. In addition to that, we even improved quality on the source domain data.