A double branches network with IoU-based loss weight to get more accurate deformable pooling feature for angle sensitive ships

Abstract

Our previous work used two deformable pooling to get different feature regions for classification and regression to alleviate the spatial misalignment problem of the rotating ships based on a three-stage framework in remote sensing image. In this paper, we find that when the IoU threshold of the third stage is 0.55, the model gets the best average precision (AP). Besides, the deformable pooling is not suitable for negative samples and the effect of deformable pooling for positive samples will gradually weaken as the Intersection-over-Union (IoU) of the sample increase. We propose a complementary branch with IoU-based loss weight to enhance the training of the pooling feature. The more accurate the pooling feature, the more accurate the deformable pooling feature. Furthermore, we add an angle offset learning module to the regression task of the complementary branch to enhance the ability of pooling feature to learn angle offset.