SRA-YOLO: Spatial Resolution Adaptive YOLO for Semi-supervised Cross-Domain Aerial Object Detection

Junhao Huang, Jian Xue, Yuqiu Li, Hao Wu, Ke Lu

Abstract: This study introduces SRA-YOLO, a novel framework tailored for Semi-Supervised Cross-Domain Aerial Object Detection, leveraging the robust YOLOv5 architecture. Aimed at addressing the challenges of spatial resolution variances and data scarcity in aerial imagery, SRA-YOLO employs an innovative Teacher-Student strategy integrating strategic knowledge distillation to utilize both labeled and unlabeled data effectively. Our approach stands out by introducing adaptive training data generation techniques, specifically Adaptive Zoom-In and Zoom-Out methods, to counteract domain discrepancies and align Ground Sample Distance (GSD) across diverse aerial conditions. Through extensive experiments on benchmark datasets, notably DOTA-v1.5, DOTA-v2.0 and xView, our method demonstrates superior adaptability and performance, setting a new baseline for aerial object detection in semi-supervised and cross-domain scenarios.