Reliable Pseudo-Labeling and Confusion Calibration for Foggy-Scene Semantic Segmentation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Driving Systems

Background:

• Semantic segmentation is vital for autonomous driving, but foggy conditions pose significant challenges.
• Acquiring annotated foggy data is expensive, limiting model training.
• Current unsupervised domain adaptation methods struggle with class confusion and unreliable pseudo-labels in foggy scenes.

Purpose of the Study:

• To develop a robust framework for semantic segmentation in foggy environments.
• To address the dual challenges of unreliable supervision signals and reduced class discriminability.
• To improve the accuracy and reliability of autonomous driving systems operating in adverse weather.

Main Methods:

• Proposed a reliable pseudo-labeling and confusion calibration framework (RPCC).
• Introduced dynamic energy-guided pseudo-labeling (DEPL) to model prediction reliability using energy scores.
• Implemented a reliable-region class confusion calibration (RCC) module to calibrate semantic relationships and suppress class confusion.

Main Results:

• The RPCC framework significantly improved the reliability of target-domain supervision signals.
• The RCC module effectively suppressed class confusion and enhanced semantic boundary clarity.
• Experiments showed RPCC outperformed existing methods on real-world foggy-scene datasets.

Conclusions:

• The proposed RPCC framework offers a superior solution for semantic segmentation in foggy conditions.
• RPCC effectively tackles the challenges of unreliable pseudo-labels and class confusion.
• The method demonstrates strong generalization capabilities across various adverse weather conditions.