ROSD: Refined Oriented Staged Detector for Object Detection in Aerial Image

Abstract

Object detection in aerial image is a challenging task. Although many advanced methods based on the convolutional neural network were popular in natural scenes, the progress in aerial images is not so smooth. Unlike natural scenes, objects in aerial images have the characteristics of arbitrary orientation, densely distribution, and large scale variation, which leads to a series of problems such as feature misalignment, missed detection, and poor detection of large aspect ratio objects. In this paper, a Refined Oriented Staged Detector (ROSD) with the combination of refined horizontal detector and rotated detector is proposed to address these problems. In our refined horizontal detector, a multi-orientation Region of Interest (RoI) Align and Orientation Attention Module (OAM) are adopted to make use of the orientation information for obtaining the orientation-sensitive features in regression branch and produce the orientation-invariant features in classification branch. Considering the feature misalignment between horizontal and rotated anchors in rotated detector, Deform Inception Module (DIM) is proposed to deal with the geometric deformation problem caused by the location changes. Besides, we propose an aspect ratio guided loss which consists of a smooth L1 loss and an angular offset penalty loss to improve the detection performance of large aspect ratio objects. Comparison experiments on two public aerial images datasets (i.e., DOTA and HRSC2016) demonstrate that our method can achieve a competitive performance.