Oriented Ship Detection Based on Expansion Deformation Rotation Representation

Existing mainstream two-stage oriented detection methods generate oriented proposals based on horizontal proposals as the circumscribed rectangle. Utilizing such generating strategy cannot obtain accurate oriented proposals due to the shape arbitrariness of ship targets. This paper proposes an effective and simple oriented proposal generating strategy, termed Expansion Deformation Rotation Representation (EDRR). EDRR introduces two expansion components that allow the oriented proposal has the potential to break through the constraints of the horizontal proposal. Then, a novel two-stage oriented ship detection framework named OSDet is proposed based on the EDRR strategy. To be specific, in the first stage, we use ReResNet-ReFPN to extract rotation invariant features and oriented RPN with EDRR strategy to generate accurate oriented candidates. Then, the oriented R-CNN head is utilized to refine oriented Rol features and complete the regression and recognition procedures in the second stage of the proposed OSDet. Extensive experiments on two challenging optical remote sensing datasets HRSC2016-MS and DOTA demonstrate that the proposed method can achieve competitive performance on the oriented ship detection task and have strong generalization capability on the aerial object detection task. Comparison with state-of-the-art methods indicate that the proposed OSDet achieves the best oriented ship/object detection performance, i.e., 77.50% and 77.14% mAP on HRSC2016-MS and DOTA datasets, respectively.