SemI2I: Semantically Consistent Image-to-Image Translation for Domain Adaptation of Remote Sensing Data

Although convolutional neural networks have been proven to be an effective\ntool to generate high quality maps from remote sensing images, their\nperformance significantly deteriorates when there exists a large domain shift\nbetween training and test data. To address this issue, we propose a new data\naugmentation approach that transfers the style of test data to training data\nusing generative adversarial networks. Our semantic segmentation framework\nconsists in first training a U-net from the real training data and then\nfine-tuning it on the test stylized fake training data generated by the\nproposed approach. Our experimental results prove that our framework\noutperforms the existing domain adaptation methods.\n