Deep learning of boundary representation from point clouds

Project Leader: Kourosh Khoshelham
Primary Contact: Kourosh Khoshelham (k.khoshelham@unimelb.edu.au)
Keywords: deep learning; geomatics; machine learning; spatial information
Disciplines: Infrastructure Engineering
Domains:

Existing methods for automated 3D reconstruction of man-made objects from point clouds mostly generate mesh models. Learning to generate boundary representation (B-rep) models from point clouds requires plenty of data which has not been available before. In recent years public datasets of building information models have become available which makes it possible to generate unlimited (synthetic) training data by using lidar simulators. The aim of this research is to explore the feasibility of deep learning from such data to generate B-rep models from point clouds.