Research on Virtual Data Set Generation for Ship Target Recognition at Sea

Deep learning provides a feasible and effective method for ship target recognition and state estimation. However, it takes a lot of manpower and financial resources to collect and label images from real sea scenes. In addition, when it is necessary to establish a data set of ships at sea with specific characteristics, such as specifying camera position, angle, altitude, different weather and sea conditions, different time, etc., it is even more difficult to obtain. Therefore, this paper proposes a method based on Unity3D to automatically generate a VirtualShip data set with accurate labeling information, and constructs a virtual data set named Virtual Ship, which is used for various computer vision recognition and tracking tasks. The VirtualShip data set is tested and verified by using YOLOv5 detector on Singaore Maritime Date (SMD), and it is calculated that mAPP0.5 can reach up to 96.0%. At the same time, by adding VirtualShip virtual data set to SMD, the value of mAP0.5:0.95 has been improved. From the experimental results, it is feasible to use the generated virtual data set for training and testing the object detector in the task of ship target recognition and tracking at sea, and it can improve the performance of the model to some extent.