A Self-Supervised Pretraining Framework for Context-Aware Building Edge Extraction From 3-D Point Clouds

Abstract

Building edge points, as essential geometric features, are crucial for advancing smart city initiatives and ensuring the precise reconstruction of 3-D structures. However, existing methods struggle to effectively design point-to-edge distance constraints for accurate building edge point identification. In this letter, we propose a novel self-supervised learning (SSL)-based pretraining framework that integrates an innovative edge point identification loss function for extracting building edge points. Specifically, we use an SSL-based feature extractor, leveraging a masked autoencoder to generate pointwise features from the input building point clouds. These features are subsequently processed by the proposed edge point identification module, which optimizes three key distance-based loss functions: the distance between any input point and its nearest edge, the distance between candidate edge points and the projection of the input point, and the distance between candidate edge points and the edges themselves. The proposed framework demonstrates superior performance in edge point extraction across both partial and complete datasets, outperforming existing methods in edge point identification.