Training semantic segmentation model relies on pixel-wise annotated images, obtaining such dense annotations is notoriously time-consuming and costly. This paper presents a data curation pipeline to facilitate cost-effective construction of semantic segmentation datasets by selecting a representative subset from a vast pool of unlabeled images for annotation. Our approach uses a distribution-preserving selection mechanism based on optimal transport to align the subset’s content with a validated reference set, ensuring that the curated images collectively cover the diverse semantic classes of the target domain. Experiments on segmentation benchmark demonstrate that training models on data curated by our method yields better segmentation performance than training on randomly selected or conventional nearest-neighbor selected subsets, particularly for under-represented object classes.