Arbitrary-Oriented Dense Object Detection in Remote Sensing Imagery

Abstract

Automatic object detection in remote sensing images is of significant importance with widespread practical applications. However, complex backgrounds, small size and dense arrangement of objects, as well as the various orientations of the target pose great challenges for current object detection algorithms. In this paper, an arbitrary-oriented dense object detection network is proposed to predict the object area using oriented bounding boxes. Firstly, we present a method to predict the object angle according to the features in the proposal, which does not increase computation costs by utilizing weight sharing. Then, a bound conversion algorithm is built to generate the oriented bounding box of an object according to the result of axis-aligned horizontal box and predicted angle information. In addition, we employ a two-stage NMS algorithm to reduce the omission ratio for dense objects by introducing oriented boxes to compute overlapping ratio. Detailed evaluations on the DOTA dataset demonstrate the effectiveness of the proposed method.