Condition-Invariant Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Robust visual perception is crucial for autonomous systems like cars and robots.
• Existing feature-level adaptation methods struggle with condition-level adaptation, often underperforming simpler pixel-level stylization techniques.

Purpose of the Study:

• To develop a novel feature-level adaptation method that leverages stylization for improved condition-level adaptation in semantic segmentation.
• To enhance the robustness of semantic segmentation networks across diverse visual conditions.

Main Methods:

• Proposed Condition-Invariant Semantic Segmentation (CISS) method, which aligns internal network features from original and stylized images using a feature invariance loss.
• Implemented CISS on a state-of-the-art domain adaptation architecture.
• Encouraged encoders to extract style-invariant features, allowing decoders to focus on semantic parsing.

Main Results:

• Achieved state-of-the-art results on the Cityscapes Dark Zurich benchmark (daytime-to-nighttime adaptation).
• Secured second-best performance on the Cityscapes ACDC benchmark (normal-to-adverse adaptation).
• Demonstrated strong generalization to unseen domains like BDD100K-night and ACDC-night.

Conclusions:

• CISS effectively improves condition-level adaptation in semantic segmentation networks.
• The proposed feature invariance loss enables robust perception across varied visual environments.
• CISS represents a significant advancement for real-world autonomous system perception.