CornerRegNet: Building Segmentation from Overhead Imagery Using Oriented Corners in Deep Networks

The increasing demand for high-resolution maps across various applications has driven the need for building segmentation vectors from overhead imagery. Nevertheless, current deep neural networks commonly output raster data, necessitating extensive post-processing and often compromising the fidelity, regularity, and simplicity of buildings. This paper proposes a novel deep convolutional neural network, CornerRegNet, which directly extracts delineated building polygons from images. Specifically, a deep model is designed to predict building footprint masks, corners, and corresponding orientation vectors pointing to adjacent corners. Then, an initial polygon is reconstructed using the predictions, and a graph convolution network repeatedly refines it using both semantic and geometric features. Initialized by the predicted corners, the output polygon is inherently simple, and the geometric information in oriented corners leads to more regular and accurate results. Compared to state-of-the-art methods, our approach demonstrates a highly competitive performance on SpaceNet Vegas and CrowdAI datasets.