Semantically Adaptive Image-to-image Translation for Domain Adaptation of Semantic Segmentation

Domain shift is a very challenging problem for semantic segmentation. Any\nmodel can be easily trained on synthetic data, where images and labels are\nartificially generated, but it will perform poorly when deployed on real\nenvironments. In this paper, we address the problem of domain adaptation for\nsemantic segmentation of street scenes. Many state-of-the-art approaches focus\non translating the source image while imposing that the result should be\nsemantically consistent with the input. However, we advocate that the image\nsemantics can also be exploited to guide the translation algorithm. To this\nend, we rethink the generative model to enforce this assumption and strengthen\nthe connection between pixel-level and feature-level domain alignment. We\nconduct extensive experiments by training common semantic segmentation models\nwith our method and show that the results we obtain on the synthetic-to-real\nbenchmarks surpass the state-of-the-art.\n