Computer Vision News - July 2021

Presentation 34 Best of CVPR 2021 BRepNet: A Topological Message Passing System for Solid Models Joe Lambourne is a Senior AI Researcher at Autodesk Research with the Autodesk AI Lab. His paper proposes a neural network architecture designed to operate directly on B-rep data structures. He speaks to us ahead of his presentation today. I begin by telling Joe what an honor it is to speak to him, as Autodesk played such a defining role at the start of my career in the 1980 s. Back then, I was selling PC hardware and software in Paris, and AutoCAD was one of my products! “ That’s wonderful. Well, obviously things have moved on a bit since then! ” Joe laughs. “ Now, we have products like Fusion and Inventor , which are full 3D CAD modelers . They T all create models in the B-rep solid model format, which is the de facto standard for defining 3D geometry in industrial CAD. ” Until recently, machine learning has focused on representations like point clouds and triangle meshes, which lose a lot of information that is present in B -rep models. With the availability of open-source CAD modelling kernels like Open Cascade , which give people the ability to read B -rep models in STEP format, Joe and his team have picked the perfect time to exploit this gap. The BRepNet architecture is motivated by the realization that graph neural networks do not have any concept of ordering of one node around other nodes. They use symmetric aggregation functions to combine and aggregate the messages which are coming from neighboring nodes at each step. With the B -rep data structure and with solid models in general, the boundary representation is defining a manifold. It knows the ordering of any neighboring entity around a particular directed edge in the model. BRepNet: A Topological Message Passing System for Solid Mo els Joe Lambourne is a Senior AI Researcher at Autodesk Research with the Autodesk AI Lab. His paper proposes a neural network architecture designed to operate directly on B-rep data structures. He speaks to us ahead of his presentation today. I begin by telling Joe what an honor it is to speak to him, as Autodesk played such a defining role at the start of my career in the 1980 s. Back then, I was selling PC hardware and software in Paris, and AutoCAD was one of my products! “ That’s wonderful. Well, obviously things have moved on a bit since then! ” Joe laughs. “ Now, we have products like Fusion and Inventor , which are full 3D CAD modelers . They all create models in the B-rep solid model format, which is the de facto standard for defining 3D geometry in industrial CAD. ” Until recently, machine learning has focused on representations like point clouds and triangle meshes, which lose a lot of information that is present in B -rep models. With the availability of open-source CAD modelling kernels like Open Cascade , which give people the ability to read B -rep models in STEP format, Joe and his team have picked the perfect time to exploit this gap. The BRepNet architecture is motivated by the realization that graph neural networks do not have any concept of ordering of one node around other nodes. They use symmetric aggregation functions to combine and aggregate the messages which are coming from neighboring nodes at each step. With the B -rep data structure and with solid models in general, the boundary representation is defining a manifold. It knows the ord ring of any neighboring entity around a particular directed edge in the model.