Building detection in high spatial resolution remote sensing imagery with the U-Rotation Detection Network

Abstract

Building detection in high spatial resolution optical remote sensing images is important for city planning, navigation, population estimation and many other applications. Although many methods have been proposed, building detection is still a challenging problem due to complex scenes and small or arbitrarily orientated buildings. Moreover, most algorithms detect rotated buildings with horizontal bounding boxes leading to many background pixels being preserved in the final detection, which is not beneficial for post-processing. To address these problems, we present the U-Rotation Detection Network (U-RDN), which can effectively detect buildings with arbitrarily orientated detection bounding boxes. First, the U-Rotation Region Proposal Network (U-RRPN) is proposed to generate rotated proposals through rotated anchors. Then, a Rotation Fast-Region Convolutional Neural Network (RFast-RCNN) is performed, which extracts fixed-size features from rotated proposals and utilizes them to obtain fine-detections. For extracting fixed-size features from rotated proposals, we propose Auto Mask Region-Of-Interest Align (AM-ROI Align). The AM-ROI Align not only reduces abundant noise but also preserves the proper information of an object in ROI. Experimental results using the public building dataset, SpaceNet, show that our method can detect buildings with skewed bounding boxes and has a state-of-the-art performance compared with other algorithms.