Enhancing point cloud semantic segmentation of building interiors through diffusion-based scene-level synthesis
Hongzhe Yue, Qian Wang, Lijie Huang, Mingyu Zhang
Abstract
Synthetic point clouds have significant potential to enhance deep learning (DL)-based semantic segmentation, thereby facilitating the reconstruction of building interior scenes. However, current methods often rely on pre-built models for sampling, which limits their applicability when such models are unavailable. This paper proposes a Diffusion-Based Scene-Level Point Cloud Synthesis (DS-PCS) method, capable of generating unlimited synthetic point clouds solely through text prompts. A total of 34 sets of comparison experiments were conducted to assess the semantic segmentation performance across different algorithms, synthetic point cloud generation approaches, and training datasets. The results show that: (1) DS-PCS generated synthetic point cloud scenes in diverse styles, colors, and types; (2) incorporating these synthetic point clouds into training datasets alongside real point clouds significantly improved semantic segmentation accuracy on the S3DIS dataset; and (3) using mixed point cloud datasets further enhanced segmentation accuracy while substantially reducing the time required for preparing training data.