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

作者：

Highlights：

• The first large-scale fine-grained building facade 3D point clouds dataset benchmark for semantic segmentation, the dataset has some proprietary characteristics.

• An enhanced position encoding block, which aggregates different spatial information to learn more local geometric structure information.

• The Dual Local Attention (DLA) module consists of two blocks including the self-attention block and the attentive pooling block.

• The proposed DLA-Net presents decent performance under the building facade dataset.

摘要

•The first large-scale fine-grained building facade 3D point clouds dataset benchmark for semantic segmentation, the dataset has some proprietary characteristics.•An enhanced position encoding block, which aggregates different spatial information to learn more local geometric structure information.•The Dual Local Attention (DLA) module consists of two blocks including the self-attention block and the attentive pooling block.•The proposed DLA-Net presents decent performance under the building facade dataset.

论文关键词：Semantic segmentation,Building facade,Self-attention,Attentive pooling,DLA-Net

论文评审过程：Received 10 May 2021, Revised 11 September 2021, Accepted 14 October 2021, Available online 19 October 2021, Version of Record 28 October 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.108372