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

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Domain Generalized Semantic Segmentation (DGSS) models struggle with style variations (texture, illumination) between training and unseen domains.
• Existing style transfer methods in DGSS often cause over-stylization, leading to semantic information loss and reduced generalization.

Purpose of the Study:

• To develop a novel approach, Style-Aware Dynamic Style Transfer (SADST), to enhance DGSS model generalization.
• To address the limitations of fixed or random style transfer strategies in DGSS.

Main Methods:

• Style Extraction Block (SEB) to extract style information from low-level features while preserving semantic cues.
• Dynamic Style Transfer Module (DSTM) to predict style transfer intensity based on original and stylized features.
• Style-Semantic Consistency Loss to ensure style-invariant representations by maintaining consistent segmentation results across stylized features.

Main Results:

• SADST significantly improves generalization performance in DGSS.
• The proposed method outperforms current state-of-the-art DGSS techniques.
• Experiments demonstrate the effectiveness of SADST in handling style discrepancies.

Conclusions:

• SADST offers a more effective approach to domain generalization in semantic segmentation.
• The dynamic and style-aware strategy preserves semantic information, leading to better performance on unseen domains.
• The method provides a robust solution for real-world applications with varying visual styles.