YOLO Based Object Detection in SAR Images: Embedded Implementation and Performance Analysis
12th International Conference on Control, Decision and Information Technologies, CoDIT 2026, Bari, Italy, 13 - 16 July 2026, pp.3419-3424, (Full Text)
- Publication Type: Conference Paper / Full Text
- Doi Number: 10.1109/codit70676.2026.11631293
- City: Bari
- Country: Italy
- Page Numbers: pp.3419-3424
- Keywords: Edge Computing, Neural Networks, Object Detection, Oriented Bounding Box (OBB), Quantization, Synthetic Aperture Radar (SAR), YOLO
- Yıldız Technical University Affiliated: Yes
Abstract
Object detection using synthetic aperture radar (SAR) imagery is a significant challenge. This paper presents a comparative analysis of YOLO26x, YOLO26x-OBB, and YOLOv8x-OBB models using the FAIR-CSAR dataset for object detection in synthetic aperture radar (SAR) imagery. While many studies evaluate SAR object detection models in offline workstation environments, their performance across different computational platforms remains largely unexplored. To address this gap, the models are deployed and evaluated on multiple platforms, including the Jetson AGX Orin, NVIDIA RTX A4000, and NVIDIA RTX PRO 6000 Blackwell Max-Q. The study also focuses on the performance of the oriented bounding box (OBB) detection, which is essential to accurately localize objects in SAR imagery. In addition to detection accuracy, the performance of the real-time inference on these platforms is analyzed. Experimental results highlight the trade-off between detection accuracy and inference latency across different computational systems, demonstrating the feasibility of deploying advanced SAR object detection models on both edge and high-performance platforms.