Uncertainty-Driven YOLO for Object Detection in Adverse Weather

Object detection in adverse weather conditions, such as fog and rain, is challenging due to degraded visual information. This paper addresses the critical issue of uncertainty quantification for robust object detection in such conditions. Existing methods, primarily focused on image enhancement or domain adaptation, often overlook aleatoric and epistemic uncertainties arising from noisy, weather-degraded data. We propose an Uncertainty-Driven Detection Framework (UDDF) that incorporates Uncertainty Quantification Networks (UQNs) – using reparameterization for aleatoric uncertainty – and Monte Carlo Dropout (MC-Dropout) for epistemic uncertainty. The UGDIP module, integrating UQNs, is seamlessly combined with a YOLOv8 detector, forming an end-to-end trainable architecture within the UDDF. Comprehensive experiments on the VOC-Foggy benchmark and the real-world RTTS dataset demonstrate that UDDF achieves state-of-the-art performance. Specifically, on the VOC-Foggy test set, UDDF achieves 81.21% mAP@0.5 (V_n_ts) and 83.10% mAP@0.5:0.95 (V_F_t), surpassing YOLOv8n by 5.63% and 8.69%, respectively, and other methods like IA-YOLO and GDIP-YOLO. On the RTTS dataset, UDDF achieves a mAP of 46.36, outperforming the best baseline (IA-YOLO) by a significant margin. The framework provides a principled approach for robust object detection, demonstrating that explicitly modeling uncertainty significantly enhances the reliability of perception systems in challenging environments.