Pre-Training of Auto-Generated Synthetic 3D Point Cloud Segmentation for Outdoor Scenes

Acquiring and annotating large datasets for the segmentation of outdoor 3D point clouds observed by airborne light detection and ranging (LiDAR) is resource-intensive and fraught with privacy concerns, limiting the availability of labeled training data. Pre-trained models can help to some extent but their effectiveness hinges on large datasets, and self-supervised learning faces data scarcity challenges. We propose a formula-driven auto-generated terrain and shape point cloud dataset for 3D point cloud segmentation tasks. Our synthetic dataset was created from diverse 3D models with variations in polygon types and shape similarity, and provides a high-quality pre-training alternative to existing datasets. Experiments reveal that models pretrained on our synthetic data outperform those trained from scratch and rival existing self-supervised learning methods. Our synthetic data aims to supplement the 3D point clouds observed by airborne LiDAR segmentation models and tackle the challenge of limited data availability in this field.