This paper focuses on the challenging problem of unsupervised domain adaptation of synthetic data for the semantic segmentation task of autonomous driving scenes. It is motivated by the generative adversarial methods that apply image-to-image translation by learning a mapping between the source and target domains. Fully supervised training of deep models for semantic segmentation do not generalize well to unseen target data. By applying domain adaptation, a model can be fit that generalizes to the target domain. Previous work has shown that combining generative adversarial networks with cycle consistency is effective for mapping images between domains, which can then be used to train a domain invariant semantic segmentation model. However, this requires additional networks to implement the cycle-consistency constraint. This paper proposes replacing this with a more efficient contrastive objective for the semantic segmentation task. By reducing the training time and computational resources, more complex end-to-end domain adaptation architectures may be used.