Energy-Based Domain Adaptation Without Intermediate Domain Dataset for Foggy Scene Segmentation

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Machine Learning

Background:

• Robust semantic segmentation in adverse weather, particularly dense fog, is critical for autonomous driving safety.
• Current methods often rely on intermediate synthetic datasets and multi-stage training, limiting real-world practicality.
• The problem of overconfident pseudo-labels in self-training for domain adaptation in foggy scenes is underexplored.

Purpose of the Study:

• To propose a novel framework, DAEN, for direct adaptation of segmentation models to real foggy scenes without intermediate datasets or multi-stage training.
• To address the overconfidence issue in pseudo-label generation during self-training for foggy scene adaptation.
• To enhance the robustness and generalization of autonomous driving perception systems in adverse weather.

Main Methods:

• Developed DAEN (Directly Adapts without additional datasets or multi-stage training) framework.
• Integrated a High-order Style Matching (HSM) module to align feature statistics between clear and foggy domains implicitly.
• Introduced Energy Score-based Pseudo-Labeling (ESPL) to generate reliable pseudo-labels by mitigating overconfidence.

Main Results:

• DAEN achieved state-of-the-art segmentation performance on three real-world foggy scene datasets.
• The HSM module effectively matched high-order statistics, enabling implicit learning of fog distributions.
• ESPL successfully mitigated pseudo-label overconfidence, improving model bias and class representation.

Conclusions:

• DAEN offers a practical and efficient solution for adapting autonomous driving perception to foggy conditions.
• The proposed method demonstrates superior performance and generalization capabilities compared to existing approaches.
• This work advances self-training techniques for domain adaptation in challenging environmental conditions.