StandardGAN: Multi-source Domain Adaptation for Semantic Segmentation of Very High Resolution Satellite Images by Data Standardization

Domain adaptation for semantic segmentation has recently been actively\nstudied to increase the generalization capabilities of deep learning models.\nThe vast majority of the domain adaptation methods tackle single-source case,\nwhere the model trained on a single source domain is adapted to a target\ndomain. However, these methods have limited practical real world applications,\nsince usually one has multiple source domains with different data\ndistributions. In this work, we deal with the multi-source domain adaptation\nproblem. Our method, namely StandardGAN, standardizes each source and target\ndomains so that all the data have similar data distributions. We then use the\nstandardized source domains to train a classifier and segment the standardized\ntarget domain. We conduct extensive experiments on two remote sensing data\nsets, in which the first one consists of multiple cities from a single country,\nand the other one contains multiple cities from different countries. Our\nexperimental results show that the standardized data generated by StandardGAN\nallow the classifiers to generate significantly better segmentation.\n