SABDR: Bidirectional Dynamic Domain Adaptation with Style Alignment for Small Object Detection Under Adverse Weather.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Small object detection is hindered by adverse weather, causing domain shifts and obscuring targets.
• Existing methods often rely on weather restoration or global feature alignment, which are insufficient for cross-weather adaptation.

Purpose of the Study:

• To develop a novel approach, SABDR, for effective cross-weather small object detection.
• To address challenges posed by weather-induced domain shifts and sparse visual cues.

Main Methods:

• Introduced the Bidirectional Dynamic Domain Adaptation Network (BiDDC-Net) for domain translation and dynamic receptive field adjustment.
• Utilized the Style-Aware Domain Adaptation Module (AIFI-DA) for feature statistics modulation and enhancing small-object channels.
• Employed Style-Direction Alignment (SDA) as a regularizer for style-direction consistency.

Main Results:

• Achieved 47.7 mAP50 on Cityscapes→Foggy Cityscapes.
• Obtained high mAP50 scores on Foggy (96.0%/96.8%), Rainy (66.7%/77.1%), and Snowy (95.0%/95.6%) MOT-Fly datasets using YOLOv5s/YOLO26.
• Demonstrated significant improvements in cross-weather small object detection.

Conclusions:

• SABDR effectively adapts small object detection models to adverse weather conditions.
• The proposed methods, BiDDC-Net and AIFI-DA, show strong performance in challenging cross-weather scenarios.
• SABDR offers a robust solution for real-world applications requiring reliable small object detection in various weather.