AOGC: Anchor-Free Oriented Object Detection Based on Gaussian Centerness

Oriented object detection is a challenging task in scene text and remote sensing image analysis, it has attracted extensive attention due to the development of deep learning recent years. Currently, mainstream oriented detectors are anchor-based methods. These methods increase computational load network cause large amount anchor box redundancy. In order address this issue, we proposed an anchor-free method based on Gaussian centerness (AOGC), which single-stage method. Our uses contextual FPN (CAFPN) obtain information target. Then, designed label assignment for objects, can select positive samples with higher quality suitable aspect ratio targets. Finally, developed kernel-based branch that effectively determine significance different anchors. AOGC achieved mAP 74.30% DOTA-1.0 datasets 89.80% HRSC2016 datasets, respectively. experimental results show exhibits superior performance other achieves similar two-stage