Source Domain Subset Sampling for Semi-Supervised Domain Adaptation in Semantic Segmentation

In this paper, we introduce source domain subset sampling (SDSS) as a new perspective of semi-supervised adaptation. We propose adaptation by and exploiting only meaningful from data for training. Our key assumption is that the entire may contain samples are unhelpful Therefore, can benefit composed solely helpful relevant samples. The proposed method effectively subsamples full to generate small-scale subset. training time reduced, performance improved with our subsampled data. To further verify scalability method, construct dataset called Ocean Ship, which comprises 500 real 200K synthetic sample images ground-truth labels. SDSS achieved state-of-the-art when applied on GTA5 Cityscapes SYNTHIA public benchmark datasets 9.13 mIoU improvement Ship over baseline model.