Online learning for ship detection in maritime surveillance

We present an object detection system and its results for finding ships in maritime video for the purpose of ship traffic management. We compare the process of training the detection system for both offline and online learning. The system is learned in an online fashion where training samples are used in an iterative cycle and discarded after each video frame. We show that this online learning gives a performance that approaches the offline learning case and in certain parts even outperforms the offline detector. The online training has considerable practical advantages, such as much smaller memory requirements, the inherent adaptivity to changing objects and an iterative extensibility. However, due to the adaptivity, the performance of the online system can locally degrade. An interesting feature is that only a few images are sufficient to already obtain reasonable detection