TQR-Net: Tighter Quadrangle-Based Convolutional Neural Network for Dense Building Instance Localization in Remote Sensing Imagery

摘要

Building localization in remote sensing imagery (RSI) is widely applied in many geoscience and remote sensing areas. However, many existing methods cannot generate accurate building contours. In this paper, we propose an effective convolutional neural network (CNN) framework, Tighter Quadrangle Network (TQR-Net), to locate buildings with quadrangular contours in RSI. Here, TQR-Net can generate regular contours for each of building targets using a CNN branch which can predict tighter quadrangles in parallel. Then, we train and test TQR-Net on a large building dataset collected from Google Earth, and the experiment results demonstrate that the proposed method can generate high-quality building contours and significantly outperforms other CNN-based detectors.