ALIGNED FEATURE FOR VECTOR-BASED ROTATED OBJECT DETECTION

Angle-based methods have become mainstream in rotated object detection, while the vector-based method has shown advantages in solving angular periodicity. However, the vector-based method uses basic CenterNet structure, where the feature misalignment and top-feature weakening problem exist, limiting the detection performance. In this paper, we explore the structure of vector-based method and integrate feature aggregation and feature alignment into the detector, promoting final detection performance. To be specific, Semantic Feedback Feature Pyramid Network (SFFPN) and Attention-based Deformable Convolution Network (ADCN) are designed accordingly, and these two parts of sub-networks are finely embedded in the detector. We hope that our discovery and designs can make vector-based a common rotation detection method.