Benchmarking YOLO Models for Classifying Marine Objects in Variable Weather Conditions

This study benchmarks YOLOv7, YOLOv10, and YOLOv11 models for marine object detection in maritime surveillance and search-and-rescue, using the SeaDronesSee v2 dataset of aerial images. This dataset consists of annotated aerial images focusing on detecting various marine objects. We review recent progress in this field, addressing challenges like poor visibility and imbalanced datasets. The models are evaluated using mean average precision (mAP@50), F1-score, precision, and recall. Results show high accuracy for all models, with YOLOv7 slightly outperforming others. However, detecting some objects, like life-saving appliances, remains challenging. The findings support the models' potential for real-world use but indicate that further research is needed to improve generalization in diverse marine environments.