Boosting Generalization of Semantic Segmentation With Unseen Style Seeking-Based Meta-Learning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Domain generalization (DG) aims for models to perform well on unseen data domains.
• Existing methods struggle with limited training data diversity, hindering generalization.
• Single-domain DG (SDG) presents a particularly challenging scenario.

Purpose of the Study:

• To propose a novel framework, unseen style seeking-based meta-learning (USSML), for single-domain generalization.
• To enhance the diversity of training data and extend distribution boundaries.
• To improve model generalization capabilities in semantic segmentation tasks.

Main Methods:

• USSML constructs multiple plausible domains with varied styles from a single source domain.
• Domain combination is applied at global and instance levels to emulate unseen images.
• Meta-learning optimizes the model across generated domains to achieve domain invariance.

Main Results:

• USSML effectively improves model generalization on unseen domains.
• The proposed method demonstrates superiority over existing approaches in extensive experiments.
• USSML is easily integrated into existing segmentation methods with minimal computational overhead.

Conclusions:

• USSML offers a robust solution for single-domain generalization challenges.
• The framework enhances domain invariance and extends training data distribution.
• This approach significantly advances the field of domain generalization for semantic segmentation.