Unsupervised Domain Adaptive Semantic Segmentation Based on Improved DAFormer

Abstract

To overcome the intensive of manual labeling tasks at the pixel level required for semantic segmentation under traditional supervised learning, an Unsupervised Domain Adaptive for Semantic Segmentation (UDASS) method based on DAFormer improved model is proposed. This model adapted the Max Mean Discrepancy (MMD) method in the regenerated Hilbert space to help the alignment of the feature distribution, the soft paste strategy to retain the partially covered image blocks to help the model to accelerate convergence, the non-convex consistency regularization at the output level to enhance the robustness of the network, and the spatial pyramid pooling framework and the decoder with large window attention collaboration to improve its consistency. The proposed method was evaluated on the public dataset, and obtained the of 2.4% mIoU improvement in GTA5-to-Cityscapes and 1.1% mIoU in SYSTHIA-to-Cityscapes, respectively, which proved that this method was effective for DAFormer improvement.