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

Small object detection under adverse weather remains challenging due to weather-induced domain shifts and sparse visual cues of small targets. In contrast to R-YOLO/QTNet and conventional UDA methods, which mainly rely on weather-specific restoration/enhancement or global feature/magnitude alignment, SABDR explicitly targets cross-weather small object adaptation through bidirectional domain translation, degradation-aware receptive-field modeling, feature-statistics modulation, and style-direction alignment. Specifically, the Bidirectional Dynamic Domain Adaptation Network, termed BiDDC-Net, translates between source and target domains and dynamically adjusts receptive fields according to weather severity. The Style-Aware Domain Adaptation Module, termed AIFI-DA, enhances discriminative small-object channels using feature statistics. SDA is further used as a complementary training-time regularizer to encourage style-direction consistency without directly matching feature magnitudes. Experiments are conducted on Cityscapes→Foggy Cityscapes and MOT-Fly→Foggy/Rainy/Snowy MOT-Fly, including newly added rainy and snowy MOT-Fly settings, with both YOLOv5s and YOLO26 evaluated on all MOT-Fly weather conditions. SABDR achieves 47.7 mAP50 on Cityscapes→Foggy Cityscapes, and obtains 96.0%/96.8%, 66.7%/77.1%, and 95.0%/95.6% mAP50 on Foggy, Rainy, and Snowy MOT-Fly with YOLOv5s/YOLO26, respectively. The improvements on MOT-Fly are reported under a fixed single-seed setting and should therefore be interpreted as single-run empirical gains rather than statistically validated improvements. These results demonstrate its effectiveness under the evaluated fog/rain/snow cross-weather small object detection settings.