Arbitrary-Oriented Ship Detection Through Center-Head Point Extraction

Ship detection in remote sensing images plays a crucial role various applications and has drawn increasing attention recent years. However, existing arbitrary-oriented ship methods are generally developed on set of predefined rotated anchor boxes. These boxes not only lead to inaccurate angle predictions but also introduce extra hyper-parameters high computational cost. Moreover, the prior knowledge size been fully exploited by methods, which hinders improvement their accuracy. Aiming at solving above issues, this paper, we propose center-head point extraction based detector (named CHPDet) achieve images. Our CHPDet formulates ships as with head points used determine direction. And Gaussian kernel is map annotations into target heatmaps. Keypoint estimation performed find center ships. Then, regressed. The orientation-invariant model (OIM) produce feature maps. Finally, use finetune results. new dataset for multi-class fixed ground sample distance (GSD) named FGSD2021. Experimental results FGSD2021 two other widely data sets, i.e., HRSC2016, UCAS-AOD demonstrate that our achieves state-of-the-art performance can well distinguish between bow stern. Code available https://github.com/zf020114/CHPDet.