Geometric Deep Learning Dramatically Enhances Benefits of Digital Building Plans
ELIX announces the launch of Enhanced Cloud Technology, an application of geometric deep learning (GDL) to greatly increase the benefits of digital building plans.
HELIXRE now applies GDL to distinguish contents of multi-billion point clouds to allow semantic understanding of building spaces, identifying specific elements like walls, floors, ceilings, furniture, and typical office clutter. Defining and isolating these elements means no one creates more accurate building clouds faster and more cost efficiently than HELIX.
GDL is a relatively new and cutting-edge form of machine learning, overcoming limitations of convolutional neural networks that solve 2D problems like image recognition, but are not well suited to the complexity of 3D spaces. By creating superpoints, HELIX reduces the problem domain to a smaller set of points to which graph convolutions are applied. The use of graph convolutions gives more context to the superpoints allowing greater predictive accuracy in object classification.
“We saw the opportunity to improve the quality and value of building point clouds with innovation in geometric deep learning,” according to Nicolas Chaulet, Director of Applied Research and Development at HELIXRE.
Enhanced Cloud Technology from HELIXRE leverages a deep learning architecture from Loic Landrieu of IGN, the French National Geographic Institute.
“We adapted Landrieu’s architecture to scale with massive data sets generated by LiDAR scanning of huge buildings,” Chaulet explained. “Enhanced Cloud Technology uses deep learning together with graph convolutions that allow the framework to learn long-range interactions between objects.”
Comparison of traditional building point clouds to HELIX Digital Twins shows the dramatic benefits of HELIX Enhanced Cloud Technology.
“The benefits of viewing the specificity of HELIX Twins with Enhanced Cloud Technology versus traditional 3D point clouds are obvious,” Chaulet added. “The ability to better see and distinguish space contents without special training or software means that building models can be more easily understood and optimized.”