A Patch Diversity Transformer for Domain Generalized Semantic Segmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Domain generalization (DG) is crucial for deep learning models to perform reliably in unseen environments.
• Representing domain-invariant context (DIC) is a key challenge in achieving effective DG.
• Transformers show promise for learning generalized features due to their global context understanding capabilities.

Purpose of the Study:

• To propose a novel method, patch diversity Transformer (PDTrans), to enhance DG for scene segmentation.
• To improve the learning of global multidomain semantic relations for better generalization.
• To address the challenge of representing domain-invariant context effectively.

Main Methods:

• Introduced patch diversity Transformer (PDTrans) utilizing self-attention mechanisms.
• Developed patch photometric perturbation (PPP) to enhance multidomain representation in global context.
• Proposed patch statistics perturbation (PSP) to model patch feature statistics under domain shifts, encoding domain-invariant semantic features.

Main Results:

• PDTrans diversifies source domains at both patch and feature levels.
• The method effectively learns context across diverse patches.
• Experimental results show significant performance improvements of PDTrans over existing state-of-the-art DG methods.

Conclusions:

• PDTrans offers a robust approach to improve domain generalization in deep learning.
• The proposed perturbations (PPP and PSP) are effective in enhancing model generalization.
• PDTrans demonstrates superior performance in scene segmentation tasks requiring domain generalization.