A Curriculum-Style Self-Training Approach for Source-Free Semantic Segmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Source-free domain adaptation (SFDA) is crucial for privacy-preserving AI.
• Traditional SFDA methods struggle with feature alignment.
• Adapting models without source data presents unique challenges.

Purpose of the Study:

• To develop a novel approach for source-free domain adaptive semantic segmentation.
• To address limitations in feature alignment for SFDA.
• To enhance model adaptability to target domains while protecting source data.

Main Methods:

• Proposing a curriculum-style self-training framework.
• Utilizing curriculum-style entropy minimization for knowledge extraction.
• Employing complementary pseudo-labeling (positive and negative) with curriculum learning.
• Implementing an information propagation scheme to reduce intra-domain discrepancy.
• Extending the method to a black-box source model scenario.

Main Results:

• Achieved state-of-the-art performance on source-free semantic segmentation.
• Demonstrated effectiveness on both synthetic-to-real and adverse condition datasets.
• Validated the approach's robustness in challenging scenarios, including black-box settings.

Conclusions:

• The proposed curriculum-style self-training is effective for SFDA semantic segmentation.
• The method successfully adapts models without direct access to source data.
• This work advances SFDA by offering a privacy-preserving and high-performance solution.