Reconstruction of lines and planes of urban buildings with angle regularization

摘要

Three-dimensional reconstruction of line and plane structures from two images is a major task in urban building modeling. However, traditional line segment (LS) matching methods frequently produce inaccurate few LS matches, and further lead to unreliable sparse 3D line-plane reconstruction. To address these issues, this paper presents an effective line-plane reconstruction method based on angle regularization. The proposed method first performs LS matching by learning the angles between planes using convolutional neural networks (CNNs). Angle regularization is used to correct unreliable LS matches and infer progressively potential 3D LSs for unmatched ones. Then, the resulting 3D LSs and planes are globally regularized by incorporating geometric constraints, image features, and plane and angle regularity terms under a unified optimization framework. Experiments on several standard datasets demonstrate that our method has clear advantages over the state-of-the-art methods.