LAUSR

In this work, we propose a new method for real-time ship segmentation during maritime surveillance missions using aircrafts with onboard video cameras. We propose a cascade model with a detection stage followed by a segmentation stage. The detection stage selects candidate regions (bounding boxes) likely to contain ships. These bounding boxes are passed to the segmentation stage, where the ship segmentation mask is then obtained. By focusing the segmentation effort only in the image regions recommended by the previous detection stage, it is possible to improve the overall image processing time and, simultaneously, to obtain a better segmentation score than the one obtained by monolithic segmentation models that are trained in an end-to-end way. Additionally, we test the viability of using a Conditional Random Field model as final boundary refinement stage: although such model can improve the segmentation results when a full segmentation approach is used, our experiments did not show any significant improvements when using our proposed cascade model. We trained the detection and segmentation models with aerial ship images from publicly available maritime datasets. We tested the cascade model on the Airbus ship detection challenge, showing real-time performance and accurate maritime ship segmentation, comparable to state-of-the-art results.