SADST: Style-aware dynamic style transfer for domain generalized semantic segmentation.

Abstract

Domain Generalized Semantic Segmentation (DGSS) aims to generalize models trained on a single source domain to unseen target domains. The generalization performance of a model is severely hindered by significant style discrepancies between domains, such as variations in texture and illumination. While existing DGSS methods employ style transfer to alleviate these discrepancies, their fixed or random transfer strategies overlook the intricate interplay between content and style. This often leads to over-stylization, which in turn causes a loss of semantic information, consequently degrading generalization on unseen domains. To overcome this limitation, we introduce a novel approach, Style-Aware Dynamic Style Transfer (SADST). Our method consists of three key components: (1) The Style Extraction Block (SEB) extracts style information from low-level features that preserves crucial semantic cues for guiding the subsequent transfer task. (2) The Dynamic Style Transfer Module (DSTM) dynamically predicts the intensity of the style transfer by considering the style information of both the original and stylized features. (3) A Style-Semantic Consistency Loss is introduced to ensure that different stylized versions of the same feature yield identical segmentation results, compelling the model to learn style-invariant representations. Experiments conducted on several DGSS baselines show that SADST outperforms current state-of-the-art methods. The code is available at https://github.com/Sinkalex/SADST.