Enhanced three-dimensional instance segmentation using multi-feature extracting point cloud neural network

Abstract

Precise three-dimensional (3D) instance segmentation of indoor scenes plays a critical role in civil engineering, including reverse engineering, size detection, and advanced structural analysis. However, existing methods often fall short in accurately segmenting complex indoor environments due to challenges of diverse material textures, irregular object shapes, and inadequate datasets. To address these limitations, this paper introduces StructNet3D, a point cloud neural network specifically designed for instance segmentation in indoor components including ceilings, floors, and walls. StructNet3D employs a novel multi-scale 3D U-Net backbone integrated with ArchExtract, which designed to capture both global context and local structural details, enabling precise segmentation of diverse indoor environments. Compared to other methods, StructNet3D achieved an AP50 of 87.7 on the proprietary dataset and 68.6 on the S3DIS dataset, demonstrating its effectiveness in accurately segmenting and classifying major structural components within diverse indoor environments.