단일 소스 도메인 적응을 위한 Global-Local Transformation 기법 연구

초록

Data augmentation is a pivotal lever for improving robustness and generalization in deep learning and is widely adopted in computer vision. Building on YOLO 11, we propose Global–Local Transformation (GLT), an augmentation strategy for single-source domain adaptation. GLT couples a Global Transformation (GT) that perturbs image-wide statistics with a Local Transformation (LT) that stochastically alters object-centric regions, enabling simultaneous shifts in global appearance and diversification of local characteristics. We instantiate GLT for RGB→infrared (IR) transfer by training on RGB source data while targeting IR imagery at test time. Experiments show that GLT effectively narrows the domain gap and yields substantial improvements over vanilla training and commonly used augmentations, balancing global changes with local variability without incurring notable training overhead. Ablation studies further indicate that GT and LT are complementary, and their combination outperforms either component alone. Overall, GLT strengthens generalization from a single source domain to a target IR domain, demonstrating practical value when labeled target data are scarce.

키워드