Oriented Object Detection: Edge Models vs. Cloud Architectures

Abstract

Aerial object detection from UAV imagery requires models capable of localizing objects with arbitrary orientations while operating under the computational constraints of onboard processing. This work evaluates eleven oriented object detection models from the YOLO-OBB, RTMDet, RoI Transformer and Strip R-CNN families. The comparison considers both detection accuracy and deployment-oriented metrics, including latency, throughput, model size and computational complexity. The evaluation uses overlapping image patches from the DOTA v1.0 dataset. Models are compared in terms of mAP and weighted F1 using confidence thresholds selected on the calibration subset. The results show that high-capacity models such as YOLOv11x-OBB and RTMDet-l are the most suitable candidates for cloud-side inference, whereas the YOLO nano variants and RTMDet-tiny offer different trade-offs for edge-side processing. Based on these findings, a hybrid edge–cloud architecture is proposed, where lightweight onboard detectors process the UAV video stream in real time and more accurate cloud-side models are reserved for ambiguous, low-confidence or high-priority detections.