Research on Unsupervised Adaptive Semantic Segmentation based on “Template” and “Paint” Generative Model

In unsupervised domain adaptive semantic segmentation, the lack of labeled data in the target domain is a common challenge. The traditional method generates a composite image in the style of the target domain by generating the joint distribution of the source domain and the target domain by generating the model, but the problems of poor sample quality and high training cost still exist. Therefore, this paper proposes a new generation strategy, which takes the label data in the source domain as a “template” and the generated data as a “paint”, and colors different object categories in the label data in the target domain style, and generates image data with the target domain style and consistent with the source domain labels. This method can improve the performance of the segmentation model while maintaining the quality of the generated samples, and significantly reduce the consumption of training time and computing resources. Experimental results show that the proposed method has achieved significant improvement in the semantic segmentation task of the target domain, and shows good robustness and generalization ability.