Spotlight on Small-scale Ship Detection: Empowering YOLO with Advanced Techniques and a Novel Dataset

Lingya Li, Zhixing Hou, Ming Ma, Jing Xiang, Chuangxin Yuan, Guihua Xia; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 784-799

In recent years, significant advancements have been made in deep learning-based ship detection methods on the ocean surface. However, publicly available maritime datasets that include categories for small-scale ships and network frameworks optimized explicitly for small-scale ship detection on the ocean surface are still limited in availability. To address the data scarcity in small-scale ship detection, bridge the gap between small-scale ship detection and general object detection, and mitigate the impact of small objects on maritime safety, we collect a multi-scale dataset with a particular emphasis on detecting small objects on the ocean surface, named the iShip-1. Leveraging this dataset, we train the S^3Det for small-scale ship detection, which remarkably detects small-scale ships on the ocean surface. Specifically, the iShip-1 comprises 17,236 images encompassing six categories captured from multiple perspectives and under various weather conditions. Notably, the Other Ship category focuses explicitly on small-scale ships. The S^3Det is optimized for detecting small-scale ships through improved backbone and neck architecture. It employs the NWD Loss instead of the traditional IoU Loss and utilizes the Feedback Cut&Paste technique for effective data augmentation. We evaluate the performance of the S^3Det on both the Seaships and iShip-1. For small-scale ship detection, S^3Det achieved a recall rate of 68.9%, a mAP50 of 73.9%, and a mAP50:90 of 39.4%. These results indicate improvements of 5.9%, 2%, and 1.2% compared to the original YOLOv8 model, respectively. Our code and dataset are available here https://github.com/li01233/S3Det.