Deep projective prediction of building facade footprints from ALS point cloud

Abstract

The automated extraction of building facade footprints (BFFs) is a critical task in surveying and remote sensing. Existing methods primarily use mobile laser scanning point cloud as the data source, with limited methods utilizing airborne laser scanning (ALS) data. This is mainly because current methods require explicit building extraction and wall detection, and the facade points in ALS point clouds are naturally sparse and prone to incompleteness, leading to insufficient robustness in rule-based wall extraction. To address this challenge, this paper presents an end-to-end method named deep projective prediction (DPP), which directly predicts BFF masks from ALS point clouds, avoiding explicit extraction of buildings and facades, thereby simplifying the BFF extraction process. Meanwhile, we introduce a back-projective attention (BPAtt) module that guides the decoding process while performing differentiable projections, enhancing the model’s sensitivity to projected feature locations. Additionally, a sparse feature completion (SFC) strategy is proposed to alleviate the impact of point cloud sparsity on footprint mask prediction. To validate the effectiveness of the DPP and facilitate relevant future research, an ALS-based BFF dataset is established, which provides more than 3k BFF annotations. Extensive experiments demonstrate that the proposed DPP achieves promising results on the BFF extraction task. The BPAtt module and SFC strategy also promote the BFF extraction performance, particularly at the boundaries of footprints.