DLA-Net: Learning dual local attention features for semantic segmentation of large-scale building facade point clouds

Abstract

The semantic segmentation of building facades is critical for various construction applications, such as urban building reconstruction and damage assessments. As there is a lack of 3D point cloud datasets related to fine-grained building facades, in this work we construct the first large-scale point cloud benchmark dataset for building facade semantic segmentation. In terms of the characteristics of building facade dataset, the existing methods of semantic segmentation cannot fully mine the local neighborhood information of point clouds; therefore, we propose an attention module that learns Dual Local Attention features, called DLA in this paper. The proposed DLA module consists of two blocks, a self-attention block and an attentive pooling block, which both embed an enhanced position encoding block. The DLA module can be easily embedded into various network architectures for point cloud segmentation, naturally resulting in a new 3D semantic segmentation network with an encoder-decoder architecture; we called this network the DLA-Net. Extensive experimental results on our constructed building facade dataset demonstrate that the proposed DLA-Net achieves better performance than the state-of-the-art methods for semantic segmentation.