Unsupervised Foggy Scene Understanding via Self Spatial-Temporal Label Diffusion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Autonomous driving systems require robust perception in adverse weather, like fog.
• Semantic segmentation in foggy driving scenes is challenging due to data scarcity and annotation difficulties.
• Self-training is a promising approach for unsupervised domain adaptation but struggles with pseudo-label quality.

Purpose of the Study:

• To develop a novel method for improving semantic segmentation in foggy driving scenes using unsupervised domain adaptation.
• To address the limitations of sparse and inaccurate pseudo-labels in self-training strategies.
• To enhance the performance of autonomous driving perception systems under foggy conditions.

Main Methods:

• Proposed a Target-Domain driven pseudo label Diffusion (TDo-Dif) scheme leveraging local spatial similarity and temporal correspondence in image sequences.
• Utilized superpixels and optical flows to identify spatial and temporal relationships for diffusing pseudo-labels.
• Introduced local spatial similarity loss and temporal contrastive loss during model re-training to ensure feature consistency.

Main Results:

• The TDo-Dif scheme significantly improved semantic segmentation performance on foggy datasets.
• Achieved 51.92% and 53.84% mean intersection-over-union (mIoU) on Foggy Zurich and Foggy Driving datasets, respectively.
• Outperformed existing state-of-the-art unsupervised domain adaptive semantic segmentation methods.

Conclusions:

• The TDo-Dif scheme effectively densifies and refines pseudo-labels for improved unsupervised domain adaptation in foggy driving scenes.
• The method demonstrates strong performance and adaptability, even for non-sequential images.
• This research contributes to more reliable perception systems for autonomous vehicles in challenging weather.