Analyzing the Impact of Training Maritime UAV Object Detection Models on Synthetic and Real Data

Unmanned aerial vehicles (UAVs) are increasingly utilized in maritime search and rescue (SAR) enabling rapid response and wide coverage at lower costs than traditional methods. However, building accurate object detection models for SAR UAVs is challenging due to limited access to high-quality labeled maritime data. Synthetic data has emerged as a valuable supplement to real data, particularly in fields where data collection is costly or limited. This study explores synthetic data's effectiveness in training maritime SAR object detection models using the SeaDroneSee dataset and evaluates varying synthetic-to-real data ratios across three architectures: Faster R-CNN, YOLOv11, and RetinaNet. Experiments are conducted on the JARVIS high-performance computing cluster, and this paper offers a detailed analysis of the findings, examining synthetic data's practical benefits and limitations in SAR applications.