Arbitrary Oriented Ship Detection in Optical Remote Sensing Images via Partially Supervised Learning

Abstract

To more accurately locate the arbitrary orientated ships in remote sensing images, recent methods turn to perform the detection via the rotated bounding box. However, these methods require all training samples to be annotated by rotated boxes. Compared with the traditional horizontal box, annotating with such a directional box is a laborious and time-consuming work. To solve this problem, we propose a novel partially supervised ship detection method by attaching an extra rbox (rotated bounding box) regression branch as well as a weight conversion function to the typical object detection network. The parameters of predicting horizontal bounding boxes in typical object detection network can be converted into those for rotated bounding box regression through the weight conversion function. With the help of this conversion, the models can be trained on a large number of samples all of which have horizontal box annotations, but only a small fraction of which have rotated box annotations. Experimental results demonstrate the effectiveness of the proposed method.