An Improved Multi‐Task Pointwise Network for Segmentation of Building Roofs in Airborne Laser Scanning Point Clouds

Abstract

Roof plane segmentation is an essential step in the process of 3D building reconstruction from airborne laser scanning (ALS) point clouds. The existing approaches either rely on human intervention to select the appropriate input parameters for different data-sets or they are not automatic and efficient. To tackle these issues, an improved multi-task pointwise network is proposed to simultaneously segment instances (that is, individual roof planes) and semantics (that is, groups of roof planes with similar geometric shapes) in point clouds. PointNet++ is used as a backbone network to extract robust features in the first step. The features from semantics branch are then added to the instance branch to facilitate the learning of instance embeddings. After that, a feature fusion module is added to the semantics branch to acquire more discriminative features from the backbone network. To increase the accuracy of semantic predictions, fused semantic features of the points belonging to the same instance are aggregated together. Finally, a mean-shift clustering algorithm is employed on instance embeddings to produce the instance predictions. Furthermore, a new roof data-set (called RoofNTNU) is established by taking ALS point clouds as training data for automatic and more general segmentation. Experiments on the new roof data-set show that the method achieves promising segmentation results: the mean precision (mPrec) of 96.2% for the instance segmentation task and mean accuracy (mAcc) of 94.4% for the semantic segmentation task.

Learning point cloud context information based on 3D transformer for more accurate and efficient classification

“…Chen et al. (2023) improved RandLANet and used it in the task of semantic segmentation of point cloud data acquired by autonomous driving. The method adds a channel attention module to the original network, allowing it to better aggregate the geometric features of the point cloud.…”

“…In addition to the general perception tasks of the point cloud, there are a number of application-specific point-based networks. Zhang and Fan (2022) proposed a point-based roof plane extraction network for buildings. PointNet++ is used as a backbone network and the features from the semantics branch are then added to the instance branch to facilitate the learning of instance embeddings.…”

High‐resolution optical remote sensing image change detection based on dense connection and attention feature fusion network

“…Due to its powerful feature representation and non‐linear problem modelling ability, deep learning has made remarkable achievements in computer vision (CV) (Canziani et al, 2016), such as object detection (Fang, Liao, et al, 2021), video surveillance (Kaliappan et al, 2022) and image restoration (Xu et al, 2022). Excitedly, it has also been gradually expanded to the remote sensing field, including segmentation of images (Wang et al, 2021) and point clouds (Zhang & Fan, 2022), image matching (Albanwan & Qin, 2022) and image denoising (Huang et al, 2022), etc. Particularly, deep‐learning‐based CD methods have outperformed traditional CD methods by a large margin due to its powerful capacity of feature representation and high‐level automation.…”