Boundary-Based Active Domain Adaptation for Semantic Segmentation Under Adverse Conditions.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing domain adaptation semantic segmentation (DASS) methods rely on pseudo-labels, which are often noisy and biased.
• This noise and bias hinder performance improvements in DASS, especially under adverse conditions.

Purpose of the Study:

• To propose a novel boundary-based active domain adaptation (ADA) framework to address limitations of pseudo-labeling in DASS.
• To efficiently select informative low-confidence and high-confidence but misclassified samples for labeling within a limited budget.

Main Methods:

• Introduced a Ranking Weighted Feature Space Impurity (RWFSI) metric to identify low-confidence samples near decision boundaries.
• Utilized Gaussian Mixture Models (GMMs) to model domain distributions and defined an Intraclass Domain Shift Score (ICDSS) to find high-confidence but misclassified samples.

Main Results:

• The proposed ADA framework significantly enhances segmentation performance compared to existing DASS and active learning (AL) methods.
• Achieved performance comparable to fully supervised methods, demonstrating its effectiveness.

Conclusions:

• The boundary-based ADA framework offers a superior approach to DASS by intelligently selecting samples for annotation.
• This method effectively mitigates issues associated with noisy pseudo-labels and improves segmentation accuracy under challenging conditions.