Polar Transformation Based Rotation Convolution Network for Oriented Object Detection

Abstract

Recently, oriented object detection has made considerable progress in the field of remote sensing image interpretation. However, there are still some challenges for oriented objects including the dense distribution, the different shapes and the multi-scale. To put the axe in the helve, we propose a Polar Oriented Object Detector (PO <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> Det) modeled in polar coordinates. The PO <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> Det consists of the Feature Rotation Alignment Module (FRAM) and the Polar Detection Head (PDH). Firstly, the FRAM is designed to enhanced representation properties of the model to objects in the situation of dense distribution and different shapes by adjusting the sampling location of the convolution kernels to adapt to the shapes, the sizes and the orientations of the targets. Then, the PDH can obtain oriented bounding boxes by predicting two polar angles as well as one polar radius in polar coordinates, where the scales of the oriented objects can be both more clearly expressed by one polar radius than Cartesian coordinates to optimize multi-scale target detection ability of the model. Experimental results on the two commonly remote sensing datasets (i.e., DOTA and HRSC2016) proved that PO <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> Det can achieve the up-to-date performance.