Image-to-Image Translation on Defined Highlighting Regions by Semi-Supervised Semantic Segmentation

Image-to-image translations made remarkable performance in Generative Adversarial Network (GAN). While recent advances are easily generated a high-quality synthesized images, it usually remains a problem to recognize complicated scenarios. We believe that a few human annotations can greatly reduce the problems. In this paper, we propose Highlight-IT, which generates synthesized images and its corresponding pixel-level semantic segmentation. In addition, segmentation can be viewed as a strong prior and guide our framework to focus on human-defined important regions. In evaluation, we experiment with various categories of unlabeled and labeled datasets. The results show that our method achieves the quality of images of the state-of-the-art framework and also the performance of the famous semantic segmentation framework. In the end, we demonstrate the qualitative results of our work and the approaches proposed by others.