Reg: Refined Generalized Focal Loss for Road Asset Detection on Thai Highways Using Vision Models

This paper presents a novel framework for detecting and segmenting critical road assets on Thai highways using the Refined Generalized Focal Loss (REG). Integrated into state-of-the-art vision-based models, the method addresses class imbalance and challenges in localizing small, underrepresented road assets such as pavilions, pedestrian bridges, and signs. To enhance both detection and segmentation, a multi-task learning strategy is employed, optimizing REG across various tasks. REG is further refined with a spatialcontextual adjustment term to model asset distribution and a probabilistic refinement to capture uncertainty in complex environments, such as varying lighting and cluttered backgrounds. Our mathematical formulation shows that REG minimizes localization and classification errors by adaptively weighting hard-to-detect instances. Experimental results demonstrate significant performance improvements, with a mAP50 of 95.25 and an F1-score of 91.68, surpassing conventional methods. This work highlights the potential of advanced loss function refinements to improve road asset detection and segmentation, enhancing road safety and infrastructure management. The full code repository is available at https://github.com/kaopanboonyuen/REG.