GTA5-4K is a subset of GTA5, a large scale dataset for semantic segmentation. This subset is extract with the proposed Semantic Layout Matching in our paper. With this subset, the domain gap in spatial distribution between GTA5 and CityScapes is reduced, so that encourage the transferabliity of representations. Moreover, this list greatly reduces the training time (24996-> 4095), which with ample domain knowledge from GTA5. We hope this dataset could faciltate the research in domain adpatative semantic segmentation.

We will provide further comparison between model trained with GTA5 with GTA5-4K in the near future.