DA-Cal: Toward Cross-Domain Calibration in Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Unsupervised Domain Adaptation (UDA) methods improve semantic segmentation but often neglect network calibration.
• Poor network calibration leads to misaligned confidence and accuracy, posing risks in safety-critical applications.
• Soft pseudo-labels can degrade performance in cross-domain scenarios due to calibration issues.

Purpose of the Study:

• To propose DA-Cal, a novel cross-domain calibration framework for semantic segmentation.
• To address the misalignment between prediction confidence and accuracy in UDA.
• To enhance the reliability of semantic segmentation in safety-critical applications.

Main Methods:

• DA-Cal transforms target domain calibration into soft pseudo-label optimization.
• A Meta Temperature Network generates pixel-level calibration parameters.
• Bi-level optimization establishes the soft pseudo-label to UDA supervision relationship, using domain-mixing strategies.

Main Results:

• DA-Cal significantly improves target domain calibration in semantic segmentation.
• The framework seamlessly integrates with existing self-training UDA methods.
• Performance gains are achieved without introducing additional inference overhead.

Conclusions:

• DA-Cal effectively enhances network calibration quality in unsupervised domain adaptation.
• The proposed framework offers a robust solution for reliable semantic segmentation.
• DA-Cal demonstrates broad applicability across various UDA segmentation benchmarks.