3D PointNet++ Implementation for Smart Construction Management
Duration: 3 years · Repo: github.com/LitonSarker/3dPointNetImplementation
Introduction
Construction projects often lack reliable and updated BIM models, which makes manual progress tracking costly and inefficient. 3D point clouds provide an alternative by capturing as-built conditions. This project demonstrates an automation pipeline using PointNet++ to segment building elements from point cloud data, supporting real-time progress monitoring.
Target audience: researchers, industry engineers, and PhD advisors interested in computer vision and construction management.
✨ Features
- CPU-only implementation of PointNet++ (SSG) for semantic segmentation
- Supports XYZ or XYZ+RGB features
- Automatic train/validation split (80:20) from raw S3DIS dataset
- Tracks progress via OA, mAcc, mIoU metrics
- Checkpointing (
last_model.pth,best_model.pth) and logging (history.json,train_log.json) - Preprocessing scripts to convert S3DIS annotations → PLY files
🏗️ Applications in Construction
- Real-time progress monitoring from 3D point clouds
- Comparison of as-built vs BIM models (or pseudo-BIM)
- Detecting installed vs missing elements on site
- Compatible with data from LiDAR, CCTV, drones
