LAUSR

The ``shared head for classification and localization'' (sibling head) has been leading the fashion of the object detection community over the past years. However, the article by Song et al. (2020) found that there is a spatial misalignment between the two object functions of sibling head and proposed an operator called task-aware spatial disentanglement (TSD) to decouple the classification and regression from the spatial dimension by generating two disentangled proposals. But there are two problems in applying TSD to detecting rotated objects. The first problem is that the deformable pooling with the disentangled proposals in TSD is not rotation-invariant. The second problem is that the angle offset is not considered when getting the disentangled proposals. To address the above problems, we propose a rotation-invariant TSD (RITSD) method. The proposed method adopts the rotation-invariant deformable pooling and angle offset learning when getting region-of-interest (RoI) features. Considering the rotated anchor increases the computational burden and training time, we use a three-stage backbone that first utilizes horizontal proposals to obtain rotated proposals and then obtains the predicted rotated bounding boxes. Finally, detailed ablation studies on our collected dataset and HRSC2016 were performed to demonstrate the feasibility and superiority of our proposed method. From the heatmap, we find the features of near bow and stern areas are beneficial for classification and the side areas of ships are good at bounding box regression.