A One-Stage Domain Adaptation Network With Image Alignment for Unsupervised Nighttime Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving Systems

Background:

• Semantic segmentation is crucial for autonomous driving, but nighttime conditions present significant challenges due to poor illumination.
• Existing unsupervised domain adaptation (UDA) methods struggle with the large appearance differences between day and night domains.
• Scarce annotated datasets for nighttime driving further hinder model development.

Purpose of the Study:

• To develop a novel domain adaptation network, DANIA, for robust nighttime semantic image segmentation.
• To address the significant domain gap between daytime and nighttime driving scenes.
• To leverage labeled daytime data and unlabeled day-night image pairs for improved nighttime segmentation.

Main Methods:

• Proposed a multi-target adaptation framework using adversarial training across three domains: labeled daytime, unlabeled daytime, and unlabeled nighttime.
• Utilized pixel-level predictions from daytime images as pseudo-supervision for nighttime counterparts.
• Incorporated image alignment with flow estimation to refine pseudo-supervision and a re-weighting strategy to enhance small object detection.

Main Results:

• DANIA achieves state-of-the-art performance on Dark Zurich and Nighttime Driving datasets.
• The one-stage adaptation framework effectively reduces the domain shift without requiring separate pre-processing models.
• Demonstrated improved prediction accuracy, particularly for small objects in nighttime conditions.

Conclusions:

• DANIA offers a significant advancement in nighttime semantic segmentation for autonomous driving.
• The proposed approach effectively handles the challenges of poor illumination and domain shift.
• This work paves the way for more reliable perception systems in adverse nighttime conditions.